Eventos de hook — Probe ao vivo nos dois harnesses

A pergunta: quais eventos de hook do ciclo de vida realmente disparam a partir de uma sessão headless nos binários instalados, com qual payload, e qual resposta cada harness honra? As regras do Leopold só viram hooks em eventos capturados — nunca a partir de uma página de docs ou de um dump de strings. Esta página é gerada por scripts/probe-hook-events.sh a partir dos payloads verbatim que ele gravou; rode de novo nos dois binários e a página se reescreve.

A resposta em uma linha: 26 dos 33 eventos que o Claude Code documenta dispararam em 2.1.259 (Claude Code); 12 dos 12 eventos que o Codex CLI documenta dispararam em codex-cli 0.152.1. Todo evento que não disparou está registrado abaixo com o gatilho tentado.

Versões

Componente	Versão
Claude Code CLI (claude --version)	2.1.259 (Claude Code)
Codex CLI (codex --version)	codex-cli 0.152.1
Alias de modelo do Claude Code usado no probe	haiku
Modelo do Codex (dos payloads capturados)	gpt-5.6-sol
jq	jq-1.7.1-apple
Python (stubs, renderizador)	Python 3.14.6
bash (hooks, driver)	3.2.57(1)-release
SO	Darwin 27.0.0 arm64
Leopold	0.22.0 (scripts/probe-hook-events.sh, scripts/probe/dump-hook.sh)
Probe executado em (UTC)	2026-09-03T01:36:54Z

Método (hermético)

• Um projeto descartável sob mktemp -d por harness (/tmp/leopold-hookprobe.gjQnXe): com git init, um CLAUDE.md / AGENTS.md, um .leopold/PLAN.md e um PLAN.md na raiz (o controle do FileChanged), um arquivo de 120 KB para forçar compactação, um comando de projeto (.claude/commands/probecmd.md) e um arquivo de role do Codex (.codex/agents/probe-reviewer.toml).
• scripts/probe/dump-hook.sh <event> plugado em todo evento da matriz documentada de cada harness pelo writer compartilhado em extensions/lib/harness.sh — leo_wire_hooks_json no .claude/settings.json do projeto, leo_wire_hooks_toml num CODEX_HOME/config.toml temporário; nada cola JSON ou TOML à mão. O hook anexa seu stdin verbatim em <out>/<harness>/<event>.jsonl — o nome do evento é o argumento dele, então um payload sem hook_event_name ainda é arquivado sob o nome certo — e responde conforme o modo (observe / allow / deny / exit 2 / systemMessage / additionalContext / aceite de elicitation / worktree).
• Claude Code: claude -p --setting-sources project --output-format stream-json --include-hook-events em ambiente limpo. O diretório de config real é apenas lido (o macOS guarda o login no Keychain — exatamente como a run 2 de Hooks do worker SDK autenticou); um CLAUDE_CONFIG_DIR temporário é usado numa única passagem, a run não autenticada os-hooks-disparam-mesmo-assim.
• Codex CLI: codex exec --json --skip-git-repo-check --dangerously-bypass-hook-trust com stdin fechado, um CODEX_HOME temporário com uma cópia do auth.json, apagada quando o probe termina.
• Um gatilho por evento: um comando Bash que passa e um que falha; a edição do próprio .leopold/PLAN.md pela sessão e um append vindo de outro processo; um spawn via Agent / spawn_agent; um TaskCreate + TaskUpdate concluído com bloqueio por exit 2; compactação forçada (CLAUDE_CODE_AUTO_COMPACT_WINDOW / CLAUDE_AUTOCOMPACT_PCT_OVERRIDE, /compact numa sessão retomada; Codex -c model_auto_compact_token_limit); uma falha de API via ANTHROPIC_BASE_URL e via model_providers.<id>.base_url do Codex apontados para um stub em Python stdlib que responde 429, 529, 500 e 401; um pedido de permissão para uma ferramenta fora do allowlist com respostas allow e deny; uma edição do settings.json no meio da sessão; uma elicitation MCP de um stub stdio; --worktree; --init / --maintenance; /model; um control request register_repo_root; SIGINT no codex exec; um arquivo de role do Codex exercitado por spawn_agent e por codex exec com --strict-config.
• A hermeticidade é afirmada, não presumida: os nomes das entradas de ~/.claude e ~/.codex reais são fingerprintados antes e depois (a mesma checagem que scripts/test-harness-lib.sh roda) e o probe falha se qualquer um mudou. Nesta execução: ~/.claude inalterado, ~/.codex MUDOU.
• Os payloads abaixo são citados verbatim; apenas valores string com mais de 480 caracteres são cortados com um marcador explícito …[+N chars]. Os arquivos brutos sob <out>/ guardam cada byte. Dados capturados — payloads, notas de gatilho e linhas de evidência — são citados como gravados, em inglês.

Matriz — Claude Code (2.1.259 (Claude Code))

Evento	Disparou	Runs	hook_event_name	session_id	Respostas exercitadas
SessionStart	sim ×40	add-dir, additional-context, command, command-exit2, compact-auto, compact-manual, compact-manual-seed, config-change, config-change-exit2, fail-401, fail-429, fail-500, fail-529, filechanged-cwd, filechanged-nested, model-switch, model-switch-exit2, perm-allow, perm-auto, perm-deny, perm-host, posttool-block, pretool-deny, prompt-exit2, setup-init, setup-maintenance, stop-block, stop-exit2, subagent, subagent-exit2, system-message, tasks, tasks-exit2, teams, tools, unauth, worktree-agent	present	present	additionalContext, systemMessage
Setup	sim ×2	setup-init, setup-maintenance	present	present	—
UserPromptSubmit	sim ×33	add-dir, additional-context, command, compact-auto, compact-manual-seed, config-change, config-change-exit2, fail-401, fail-429, fail-500, fail-529, filechanged-cwd, filechanged-nested, perm-allow, perm-auto, perm-deny, perm-host, posttool-block, pretool-deny, prompt-exit2, setup-init, setup-maintenance, stop-block, stop-exit2, subagent, subagent-exit2, system-message, tasks, tasks-exit2, teams, tools, unauth, worktree-agent	present	present	additionalContext, exit2, systemMessage
UserPromptExpansion	sim ×2	command, command-exit2	present	present	exit2
PreToolUse	sim ×48	additional-context, compact-auto, compact-manual-seed, config-change, config-change-exit2, filechanged-cwd, filechanged-nested, perm-allow, perm-auto, perm-deny, perm-host, posttool-block, pretool-deny, subagent, subagent-exit2, system-message, tasks, tasks-exit2, tools, worktree-agent	present	present	additionalContext, permissionDecision:deny, systemMessage
PermissionRequest	sim ×2	perm-allow, perm-deny	present	present	decision:allow, decision:allow (host prompts), decision:deny
PermissionDenied	não disparou	—	—	—	—
PostToolUse	sim ×43	additional-context, compact-auto, compact-manual-seed, config-change, config-change-exit2, filechanged-cwd, filechanged-nested, perm-allow, posttool-block, subagent, subagent-exit2, system-message, tasks, tasks-exit2, tools, worktree-agent	present	present	additionalContext, decision:block, systemMessage
PostToolUseFailure	sim ×1	tools	present	present	—
PostToolBatch	sim ×48	additional-context, compact-auto, compact-manual-seed, config-change, config-change-exit2, filechanged-cwd, filechanged-nested, perm-allow, perm-auto, perm-deny, perm-host, posttool-block, pretool-deny, subagent, subagent-exit2, system-message, tasks, tasks-exit2, tools, worktree-agent	present	present	additionalContext, systemMessage
Notification	não disparou	—	—	—	—
MessageDisplay	sim ×48	add-dir, additional-context, command, compact-auto, compact-manual-seed, config-change, config-change-exit2, fail-401, fail-429, fail-500, fail-529, filechanged-cwd, filechanged-nested, perm-allow, perm-auto, perm-deny, perm-host, posttool-block, pretool-deny, setup-init, setup-maintenance, stop-block, stop-exit2, subagent, subagent-exit2, system-message, tasks, tasks-exit2, teams, tools, unauth, worktree-agent	present	present	systemMessage
SubagentStart	sim ×3	subagent, subagent-exit2, worktree-agent	present	present	—
SubagentStop	sim ×6	compact-auto, compact-manual, subagent, subagent-exit2, worktree-agent	present	present	exit2
TaskCreated	sim ×2	tasks, tasks-exit2	present	present	—
TaskCompleted	sim ×2	tasks, tasks-exit2	present	present	exit2
Stop	sim ×29	add-dir, additional-context, command, compact-auto, compact-manual-seed, config-change, config-change-exit2, filechanged-cwd, filechanged-nested, perm-allow, perm-auto, perm-deny, perm-host, posttool-block, pretool-deny, setup-init, setup-maintenance, stop-block, stop-exit2, subagent, subagent-exit2, system-message, tasks, tasks-exit2, teams, tools, worktree-agent	present	present	decision:block, exit2, systemMessage
StopFailure	sim ×5	fail-500	present	present	—
TeammateIdle	não disparou	—	—	—	—
InstructionsLoaded	sim ×38	add-dir, additional-context, command, command-exit2, compact-auto, compact-manual, compact-manual-seed, config-change, config-change-exit2, fail-401, fail-429, fail-500, fail-529, filechanged-cwd, filechanged-nested, model-switch, model-switch-exit2, perm-allow, perm-auto, perm-deny, perm-host, posttool-block, pretool-deny, prompt-exit2, setup-init, setup-maintenance, stop-block, stop-exit2, subagent, subagent-exit2, system-message, tasks, tasks-exit2, teams, tools, unauth, worktree-agent	present	present	additionalContext, systemMessage
ConfigChange	sim ×2	config-change, config-change-exit2	present	present	exit2
CwdChanged	sim ×2	filechanged-cwd, tools	present	present	systemMessage
DirectoryAdded	não disparou	—	—	—	—
FileChanged	sim ×6	filechanged-nested, tools	present	present	—
WorktreeCreate	sim ×2	worktree, worktree-agent	present	present	worktree
WorktreeRemove	não disparou	—	—	—	—
PreCompact	sim ×2	compact-auto, compact-manual	present	present	—
PostCompact	sim ×2	compact-auto, compact-manual	present	present	—
PreModelSwitch	sim ×2	model-switch, model-switch-exit2	present	present	exit2
PostModelSwitch	sim ×1	model-switch	present	present	—
Elicitation	não disparou	—	—	—	action:accept+content
ElicitationResult	não disparou	—	—	—	—
SessionEnd	sim ×34	add-dir, additional-context, command, command-exit2, compact-auto, compact-manual, compact-manual-seed, config-change, config-change-exit2, fail-429, filechanged-cwd, filechanged-nested, model-switch, model-switch-exit2, perm-allow, perm-auto, perm-deny, perm-host, posttool-block, pretool-deny, prompt-exit2, setup-init, setup-maintenance, stop-block, stop-exit2, subagent, subagent-exit2, system-message, tasks, tasks-exit2, teams, tools, worktree-agent	present	present	systemMessage

Matriz — Codex CLI (codex-cli 0.152.1)

Evento	Disparou	Runs	hook_event_name	session_id	Respostas exercitadas
SessionStart	sim ×25	additional-context, compact, exec-as-role-4, fail-401, fail-429, fail-500, fail-529, interrupt, perm-allow, perm-deny, perm-onrequest, posttool-block, pretool-deny, prompt-exit2, role-bogus, role-bogus-strict, stop-block, stop-exit2, strict-config, subagent, subagent-exit2, subagent-role, system-message, tools	present	present	--strict-config with the probe's hook block, additionalContext, systemMessage
SessionEnd	sim ×24	additional-context, compact, exec-as-role-4, fail-401, fail-429, fail-500, fail-529, interrupt, perm-allow, perm-deny, perm-onrequest, posttool-block, pretool-deny, prompt-exit2, role-bogus, role-bogus-strict, stop-block, stop-exit2, strict-config, subagent, subagent-exit2, subagent-role, system-message, tools	present	present	systemMessage
SubagentStart	sim ×3	subagent, subagent-exit2, subagent-role	present	present	codex exec -c agent_role="probe-reviewer", codex exec -c agent_type="probe-reviewer", codex exec -c agents.probe-reviewer.config_file=".codex/agents/probe-reviewer.toml", codex exec -c role="probe-reviewer", codex exec -p <role file as profile>, role file, role file with an unknown key, role file with an unknown key (--strict-config)
SubagentStop	sim ×4	subagent, subagent-exit2, subagent-role	present	present	exit2
PreToolUse	sim ×29	additional-context, compact, interrupt, perm-allow, perm-deny, perm-onrequest, posttool-block, pretool-deny, role-bogus, role-bogus-strict, subagent, subagent-exit2, subagent-role, system-message, tools	present	present	additionalContext, permissionDecision:deny, systemMessage
PermissionRequest	sim ×1	perm-deny	present	present	decision:allow, decision:allow (approval_policy=on-request), decision:deny
PostToolUse	sim ×24	additional-context, compact, perm-allow, perm-deny, perm-onrequest, posttool-block, subagent, subagent-exit2, subagent-role, system-message, tools	present	present	additionalContext, decision:block, systemMessage
PreCompact	sim ×1	compact	present	present	—
PostCompact	sim ×1	compact	present	present	—
UserPromptSubmit	sim ×24	additional-context, compact, exec-as-role-4, fail-401, fail-429, fail-500, fail-529, interrupt, perm-allow, perm-deny, perm-onrequest, posttool-block, pretool-deny, prompt-exit2, role-bogus, role-bogus-strict, stop-block, stop-exit2, strict-config, subagent, subagent-exit2, subagent-role, system-message, tools	present	present	additionalContext, exit2, systemMessage
Stop	sim ×20	additional-context, compact, exec-as-role-4, perm-allow, perm-deny, perm-onrequest, posttool-block, pretool-deny, role-bogus, role-bogus-strict, stop-block, stop-exit2, strict-config, subagent, subagent-exit2, subagent-role, system-message, tools	present	present	decision:block, exit2, systemMessage
Interrupt	sim ×1	interrupt	present	present	—

Eventos — Claude Code

SessionStart — Claude Code

Disparou 40 vez(es) na(s) run(s) add-dir, additional-context, command, command-exit2, compact-auto, compact-manual, compact-manual-seed, config-change, config-change-exit2, fail-401, fail-429, fail-500, fail-529, filechanged-cwd, filechanged-nested, model-switch, model-switch-exit2, perm-allow, perm-auto, perm-deny, perm-host, posttool-block, pretool-deny, prompt-exit2, setup-init, setup-maintenance, stop-block, stop-exit2, subagent, subagent-exit2, system-message, tasks, tasks-exit2, teams, tools, unauth, worktree-agent. hook_event_name: present; session_id: present. Chaves do payload: context_tokens, cwd, estimated_cache_write_usd, hook_event_name, model, prompt_cache_likely_expired, prompt_id, seconds_since_last_response, session_id, source, transcript_path.

Payload, verbatim (source: "startup"):

{"session_id":"777e68b4-1ad1-4f6c-9d81-64d0cae95a84","transcript_path":"/Users/jonhvmp/.claude/projects/-private-tmp-leopold-hookprobe-gjQnXe-claude-project/777e68b4-1ad1-4f6c-9d81-64d0cae95a84.jsonl","cwd":"/private/tmp/leopold-hookprobe.gjQnXe/claude-project","hook_event_name":"SessionStart","source":"startup"}

Payload, verbatim (source: "compact"):

{"session_id":"8696992c-e252-4b26-aa6e-0f09b973a8fc","transcript_path":"/Users/jonhvmp/.claude/projects/-private-tmp-leopold-hookprobe-gjQnXe-claude-project/8696992c-e252-4b26-aa6e-0f09b973a8fc.jsonl","cwd":"/private/tmp/leopold-hookprobe.gjQnXe/claude-project","prompt_id":"b1988ab3-5c85-48da-81fd-b0bd3a241365","hook_event_name":"SessionStart","source":"compact","model":"claude-haiku-4-5-20251001"}

Payload, verbatim (source: "resume"):

{"session_id":"ea4f4119-202c-4931-9721-72c83db1f7af","transcript_path":"/Users/jonhvmp/.claude/projects/-private-tmp-leopold-hookprobe-gjQnXe-claude-project/ea4f4119-202c-4931-9721-72c83db1f7af.jsonl","cwd":"/private/tmp/leopold-hookprobe.gjQnXe/claude-project","hook_event_name":"SessionStart","source":"resume","seconds_since_last_response":1,"context_tokens":42637,"prompt_cache_likely_expired":false,"estimated_cache_write_usd":0.0853}

Gatilhos tentados:

• run tools: a headless claude -p turn in a project with a CLAUDE.md that runs true, false, grep -c zzz /dev/null, an Edit of .leopold/PLAN.md and of the root PLAN.md, sleep 6 (the driver appends to both files from another process meanwhile), cat and, last, cd sub && pwd

Respostas que o harness honrou:

• systemMessage → honrada — marker PROBE_SYSMSG SessionStart appears 1 times in the stream-json output (outside hook_response echoes: 5 informational;); in the transcript: 2 PROBE_SYSMSG PostToolBatch; 2 PROBE_SYSMSG PostToolUse; 2 PROBE_SYSMSG PreToolUse; 2 PROBE_SYSMSG SessionStart; 2 PROBE_SYSMSG Stop; 1 PROBE_SYSMSG UserPromptSubmit; (run system-message)
• additionalContext → honrada — the model repeated 'PROBE_CTX SessionStart': yes; the marker is in the transcript 3 time(s) (event fired in the run: 1 times) (run additional-context)

Setup — Claude Code

Disparou 2 vez(es) na(s) run(s) setup-init, setup-maintenance. hook_event_name: present; session_id: present. Chaves do payload: cwd, hook_event_name, session_id, transcript_path, trigger.

Payload, verbatim (trigger: "init"):

{"session_id":"a16a5ec3-e3a5-44d2-8c82-b0e265ff8ea0","transcript_path":"/Users/jonhvmp/.claude/projects/-private-tmp-leopold-hookprobe-gjQnXe-claude-project/a16a5ec3-e3a5-44d2-8c82-b0e265ff8ea0.jsonl","cwd":"/private/tmp/leopold-hookprobe.gjQnXe/claude-project","hook_event_name":"Setup","trigger":"init"}

Payload, verbatim (trigger: "maintenance"):

{"session_id":"6387e875-a636-4eb4-89b5-f67903067f13","transcript_path":"/Users/jonhvmp/.claude/projects/-private-tmp-leopold-hookprobe-gjQnXe-claude-project/6387e875-a636-4eb4-89b5-f67903067f13.jsonl","cwd":"/private/tmp/leopold-hookprobe.gjQnXe/claude-project","hook_event_name":"Setup","trigger":"maintenance"}

Gatilhos tentados:

• run setup-init: claude -p --init; run setup-maintenance: claude -p --maintenance

UserPromptSubmit — Claude Code

Disparou 33 vez(es) na(s) run(s) add-dir, additional-context, command, compact-auto, compact-manual-seed, config-change, config-change-exit2, fail-401, fail-429, fail-500, fail-529, filechanged-cwd, filechanged-nested, perm-allow, perm-auto, perm-deny, perm-host, posttool-block, pretool-deny, prompt-exit2, setup-init, setup-maintenance, stop-block, stop-exit2, subagent, subagent-exit2, system-message, tasks, tasks-exit2, teams, tools, unauth, worktree-agent. hook_event_name: present; session_id: present. Chaves do payload: cwd, hook_event_name, permission_mode, prompt, prompt_id, session_id, transcript_path.

Payload, verbatim:

{"session_id":"777e68b4-1ad1-4f6c-9d81-64d0cae95a84","transcript_path":"/Users/jonhvmp/.claude/projects/-private-tmp-leopold-hookprobe-gjQnXe-claude-project/777e68b4-1ad1-4f6c-9d81-64d0cae95a84.jsonl","cwd":"/private/tmp/leopold-hookprobe.gjQnXe/claude-project","prompt_id":"4c67eb26-a782-4cfe-8201-a875dc0f7985","permission_mode":"bypassPermissions","hook_event_name":"UserPromptSubmit","prompt":"Do these steps in order with tools, one tool call per step, never two in parallel: 1) Bash: true  2) Bash: false  3) Bash: grep -c zzz /dev/null  4) Edit tool: in .leopold/PLAN.md replace 'item one' with 'item one (edited)'  5) Edit tool: in PLAN.md (project root) replace 'root item' with 'root item (edited)'  6) Bash: sleep 6  7) Bash: cat .leopold/PLAN.md  8) Bash: cd sub && pwd. Then reply with the single word DONE."}

Gatilhos tentados:

• run tools: a headless claude -p turn in a project with a CLAUDE.md that runs true, false, grep -c zzz /dev/null, an Edit of .leopold/PLAN.md and of the root PLAN.md, sleep 6 (the driver appends to both files from another process meanwhile), cat and, last, cd sub && pwd

Respostas que o harness honrou:

• exit2 → honrada — hook_response: outcome=error exit_code=2; assistant messages carrying the prompt's marker: 0; Stop fired: 0; final: UserPromptSubmit operation blocked by hook: (run prompt-exit2)
• systemMessage → honrada — marker PROBE_SYSMSG UserPromptSubmit appears 2 times in the stream-json output (outside hook_response echoes: 5 informational;); in the transcript: 2 PROBE_SYSMSG PostToolBatch; 2 PROBE_SYSMSG PostToolUse; 2 PROBE_SYSMSG PreToolUse; 2 PROBE_SYSMSG SessionStart; 2 PROBE_SYSMSG Stop; 1 PROBE_SYSMSG UserPromptSubmit; (run system-message)
• additionalContext → honrada — the model repeated 'PROBE_CTX UserPromptSubmit': no; the marker is in the transcript 2 time(s) (event fired in the run: 1 times) (run additional-context)

UserPromptExpansion — Claude Code

Disparou 2 vez(es) na(s) run(s) command, command-exit2. hook_event_name: present; session_id: present. Chaves do payload: command_args, command_name, command_source, cwd, expansion_type, hook_event_name, permission_mode, prompt, prompt_id, session_id, transcript_path.

Payload, verbatim:

{"session_id":"b2c0e499-5929-43e8-ad86-0161c31208f7","transcript_path":"/Users/jonhvmp/.claude/projects/-private-tmp-leopold-hookprobe-gjQnXe-claude-project/b2c0e499-5929-43e8-ad86-0161c31208f7.jsonl","cwd":"/private/tmp/leopold-hookprobe.gjQnXe/claude-project","prompt_id":"c76c421a-e281-4814-aa56-23155a1840da","permission_mode":"bypassPermissions","hook_event_name":"UserPromptExpansion","expansion_type":"slash_command","command_name":"probecmd","command_args":"","command_source":"projectSettings","prompt":"/probecmd"}

Gatilhos tentados:

• run command: the user prompt is the project command /probecmd (.claude/commands/probecmd.md)

Respostas que o harness honrou:

• exit2 → honrada — hook_response: outcome=error exit_code=2; the command's expansion text reached the model: no; final reply: UserPromptExpansion operation blocked by hook: (run command-exit2)

PreToolUse — Claude Code

Disparou 48 vez(es) na(s) run(s) additional-context, compact-auto, compact-manual-seed, config-change, config-change-exit2, filechanged-cwd, filechanged-nested, perm-allow, perm-auto, perm-deny, perm-host, posttool-block, pretool-deny, subagent, subagent-exit2, system-message, tasks, tasks-exit2, tools, worktree-agent. hook_event_name: present; session_id: present. Chaves do payload: agent_id, agent_type, cwd, hook_event_name, permission_mode, prompt_id, session_id, tool_input, tool_name, tool_use_id, transcript_path.

Payload, verbatim (Bash: {"command":"true","description":"Step 1: Run true command"}):

{"session_id":"777e68b4-1ad1-4f6c-9d81-64d0cae95a84","transcript_path":"/Users/jonhvmp/.claude/projects/-private-tmp-leopold-hookprobe-gjQnXe-claude-project/777e68b4-1ad1-4f6c-9d81-64d0cae95a84.jsonl","cwd":"/private/tmp/leopold-hookprobe.gjQnXe/claude-project","prompt_id":"4c67eb26-a782-4cfe-8201-a875dc0f7985","permission_mode":"bypassPermissions","hook_event_name":"PreToolUse","tool_name":"Bash","tool_input":{"command":"true","description":"Step 1: Run true command"},"tool_use_id":"toolu_012pvg9iy5XR1HxLRHTMbFL8"}

Payload, verbatim (Bash: {"command":"false","description":"Step 2: Run false command"}):

{"session_id":"777e68b4-1ad1-4f6c-9d81-64d0cae95a84","transcript_path":"/Users/jonhvmp/.claude/projects/-private-tmp-leopold-hookprobe-gjQnXe-claude-project/777e68b4-1ad1-4f6c-9d81-64d0cae95a84.jsonl","cwd":"/private/tmp/leopold-hookprobe.gjQnXe/claude-project","prompt_id":"4c67eb26-a782-4cfe-8201-a875dc0f7985","permission_mode":"bypassPermissions","hook_event_name":"PreToolUse","tool_name":"Bash","tool_input":{"command":"false","description":"Step 2: Run false command"},"tool_use_id":"toolu_01QTAduNmuev93E45QX6sgsE"}

Payload, verbatim (Bash: {"command":"grep -c zzz /dev/null","description":"Step 3: Count matches of zzz i):

{"session_id":"777e68b4-1ad1-4f6c-9d81-64d0cae95a84","transcript_path":"/Users/jonhvmp/.claude/projects/-private-tmp-leopold-hookprobe-gjQnXe-claude-project/777e68b4-1ad1-4f6c-9d81-64d0cae95a84.jsonl","cwd":"/private/tmp/leopold-hookprobe.gjQnXe/claude-project","prompt_id":"4c67eb26-a782-4cfe-8201-a875dc0f7985","permission_mode":"bypassPermissions","hook_event_name":"PreToolUse","tool_name":"Bash","tool_input":{"command":"grep -c zzz /dev/null","description":"Step 3: Count matches of zzz in /dev/null"},"tool_use_id":"toolu_01BkUCMD3BxMCBKacum25xQg"}

Payload, verbatim (Read: {"file_path":"/private/tmp/leopold-hookprobe.gjQnXe/claude-project/.leopold/PLAN):

{"session_id":"777e68b4-1ad1-4f6c-9d81-64d0cae95a84","transcript_path":"/Users/jonhvmp/.claude/projects/-private-tmp-leopold-hookprobe-gjQnXe-claude-project/777e68b4-1ad1-4f6c-9d81-64d0cae95a84.jsonl","cwd":"/private/tmp/leopold-hookprobe.gjQnXe/claude-project","prompt_id":"4c67eb26-a782-4cfe-8201-a875dc0f7985","permission_mode":"bypassPermissions","hook_event_name":"PreToolUse","tool_name":"Read","tool_input":{"file_path":"/private/tmp/leopold-hookprobe.gjQnXe/claude-project/.leopold/PLAN.md"},"tool_use_id":"toolu_01WpaqWWUfNkkeDCWHzoNWHm"}

Gatilhos tentados:

• run tools: a headless claude -p turn in a project with a CLAUDE.md that runs true, false, grep -c zzz /dev/null, an Edit of .leopold/PLAN.md and of the root PLAN.md, sleep 6 (the driver appends to both files from another process meanwhile), cat and, last, cd sub && pwd

Respostas que o harness honrou:

• observe → n/a — inside the subagent, PreToolUse carries agent_id: acb594f61d1631b05 (run subagent)
• permissionDecision:deny → honrada — PostToolUse fired 0 times after the deny (0 = the tool never ran); reason reached the model: yes (run pretool-deny)
• systemMessage → honrada — marker PROBE_SYSMSG PreToolUse appears 2 times in the stream-json output (outside hook_response echoes: 5 informational;); in the transcript: 2 PROBE_SYSMSG PostToolBatch; 2 PROBE_SYSMSG PostToolUse; 2 PROBE_SYSMSG PreToolUse; 2 PROBE_SYSMSG SessionStart; 2 PROBE_SYSMSG Stop; 1 PROBE_SYSMSG UserPromptSubmit; (run system-message)
• additionalContext → honrada — the model repeated 'PROBE_CTX PreToolUse': no; the marker is in the transcript 3 time(s) (event fired in the run: 1 times) (run additional-context)

PermissionRequest — Claude Code

Disparou 2 vez(es) na(s) run(s) perm-allow, perm-deny. hook_event_name: present; session_id: present. Chaves do payload: cwd, hook_event_name, permission_mode, permission_suggestions, prompt_id, session_id, tool_input, tool_name, transcript_path.

Payload, verbatim:

{"session_id":"12f459b8-ddba-474a-8dbb-05427ec73267","transcript_path":"/Users/jonhvmp/.claude/projects/-private-tmp-leopold-hookprobe-gjQnXe-claude-project/12f459b8-ddba-474a-8dbb-05427ec73267.jsonl","cwd":"/private/tmp/leopold-hookprobe.gjQnXe/claude-project","prompt_id":"7fabe826-67ae-42c3-9c36-d334df729b6c","permission_mode":"default","hook_event_name":"PermissionRequest","tool_name":"Bash","tool_input":{"command":"touch probe-allow.txt && echo PROBE_ALLOWED","description":"Create probe-allow.txt file and echo PROBE_ALLOWED"},"permission_suggestions":[{"type":"addDirectories","directories":["/private/tmp/leopold-hookprobe.gjQnXe/claude-project"],"destination":"session"},{"type":"setMode","mode":"acceptEdits","destination":"session"}]}

Gatilhos tentados:

• run perm-allow: default permission mode with --permission-prompts none; Bash touch (not auto-approved) needs a decision
• run perm-host: the same Bash touch with the default --permission-prompts host (a plain CLI -p has no host)

Respostas que o harness honrou:

• decision:allow → honrada — the file the command creates exists: yes; PostToolUse fired: 1 (run perm-allow)
• decision:deny → honrada — the file exists: no; message PROBE_DENY reached the model: yes; PermissionDenied fired: 0 (run perm-deny)
• decision:allow (host prompts) → não honrada — with --permission-prompts host (the default) the hook fired 0 times and the file exists: no — permission_denials in the result: 1 (run perm-host)

PermissionDenied — Claude Code

Não disparou no Claude Code 2.1.259 (Claude Code). Gatilho tentado: run perm-auto: --permission-mode auto with a Bash touch (denied, permission_denials=1); also run perm-deny: a PermissionRequest hook answering deny

PostToolUse — Claude Code

Disparou 43 vez(es) na(s) run(s) additional-context, compact-auto, compact-manual-seed, config-change, config-change-exit2, filechanged-cwd, filechanged-nested, perm-allow, posttool-block, subagent, subagent-exit2, system-message, tasks, tasks-exit2, tools, worktree-agent. hook_event_name: present; session_id: present. Chaves do payload: agent_id, agent_type, cwd, duration_ms, hook_event_name, permission_mode, prompt_id, session_id, tool_input, tool_name, tool_response, tool_use_id, transcript_path.

Payload, verbatim (Bash: {"command":"true","description":"Step 1: Run true command"}):

{"session_id":"777e68b4-1ad1-4f6c-9d81-64d0cae95a84","transcript_path":"/Users/jonhvmp/.claude/projects/-private-tmp-leopold-hookprobe-gjQnXe-claude-project/777e68b4-1ad1-4f6c-9d81-64d0cae95a84.jsonl","cwd":"/private/tmp/leopold-hookprobe.gjQnXe/claude-project","prompt_id":"4c67eb26-a782-4cfe-8201-a875dc0f7985","permission_mode":"bypassPermissions","hook_event_name":"PostToolUse","tool_name":"Bash","tool_input":{"command":"true","description":"Step 1: Run true command"},"tool_response":{"stdout":"","stderr":"","interrupted":false,"isImage":false,"noOutputExpected":false},"tool_use_id":"toolu_012pvg9iy5XR1HxLRHTMbFL8","duration_ms":90}

Payload, verbatim (Bash: {"command":"grep -c zzz /dev/null","description":"Step 3: Count matches of zzz i):

{"session_id":"777e68b4-1ad1-4f6c-9d81-64d0cae95a84","transcript_path":"/Users/jonhvmp/.claude/projects/-private-tmp-leopold-hookprobe-gjQnXe-claude-project/777e68b4-1ad1-4f6c-9d81-64d0cae95a84.jsonl","cwd":"/private/tmp/leopold-hookprobe.gjQnXe/claude-project","prompt_id":"4c67eb26-a782-4cfe-8201-a875dc0f7985","permission_mode":"bypassPermissions","hook_event_name":"PostToolUse","tool_name":"Bash","tool_input":{"command":"grep -c zzz /dev/null","description":"Step 3: Count matches of zzz in /dev/null"},"tool_response":{"stdout":"0","stderr":"","interrupted":false,"isImage":false,"returnCodeInterpretation":"No matches found","noOutputExpected":false},"tool_use_id":"toolu_01BkUCMD3BxMCBKacum25xQg","duration_ms":17}

Payload, verbatim (Read: {"file_path":"/private/tmp/leopold-hookprobe.gjQnXe/claude-project/.leopold/PLAN):

{"session_id":"777e68b4-1ad1-4f6c-9d81-64d0cae95a84","transcript_path":"/Users/jonhvmp/.claude/projects/-private-tmp-leopold-hookprobe-gjQnXe-claude-project/777e68b4-1ad1-4f6c-9d81-64d0cae95a84.jsonl","cwd":"/private/tmp/leopold-hookprobe.gjQnXe/claude-project","prompt_id":"4c67eb26-a782-4cfe-8201-a875dc0f7985","permission_mode":"bypassPermissions","hook_event_name":"PostToolUse","tool_name":"Read","tool_input":{"file_path":"/private/tmp/leopold-hookprobe.gjQnXe/claude-project/.leopold/PLAN.md"},"tool_response":{"type":"text","file":{"filePath":"/private/tmp/leopold-hookprobe.gjQnXe/claude-project/.leopold/PLAN.md","content":"# Plan\n- [ ] item one\n","numLines":3,"startLine":1,"totalLines":3}},"tool_use_id":"toolu_01WpaqWWUfNkkeDCWHzoNWHm","duration_ms":2}

Payload, verbatim (Edit: {"file_path":"/private/tmp/leopold-hookprobe.gjQnXe/claude-project/.leopold/PLAN):

{"session_id":"777e68b4-1ad1-4f6c-9d81-64d0cae95a84","transcript_path":"/Users/jonhvmp/.claude/projects/-private-tmp-leopold-hookprobe-gjQnXe-claude-project/777e68b4-1ad1-4f6c-9d81-64d0cae95a84.jsonl","cwd":"/private/tmp/leopold-hookprobe.gjQnXe/claude-project","prompt_id":"4c67eb26-a782-4cfe-8201-a875dc0f7985","permission_mode":"bypassPermissions","hook_event_name":"PostToolUse","tool_name":"Edit","tool_input":{"file_path":"/private/tmp/leopold-hookprobe.gjQnXe/claude-project/.leopold/PLAN.md","old_string":"- [ ] item one","new_string":"- [ ] item one (edited)","replace_all":false},"tool_response":{"filePath":"/private/tmp/leopold-hookprobe.gjQnXe/claude-project/.leopold/PLAN.md","oldString":"- [ ] item one","newString":"- [ ] item one (edited)","originalFile":"# Plan\n- [ ] item one\n","structuredPatch":[{"oldStart":1,"oldLines":2,"newStart":1,"newLines":2,"lines":[" # Plan","-- [ ] item one","+- [ ] item one (edited)"]}],"userModified":false,"replaceAll":false},"tool_use_id":"toolu_01F3Ro7STSqrLt9kMqMwtkZK","duration_ms":3}

Gatilhos tentados:

• run tools: a headless claude -p turn in a project with a CLAUDE.md that runs true, false, grep -c zzz /dev/null, an Edit of .leopold/PLAN.md and of the root PLAN.md, sleep 6 (the driver appends to both files from another process meanwhile), cat and, last, cd sub && pwd

Respostas que o harness honrou:

• observe → n/a — Bash true (exit 0) fired PostToolUse: yes; the failing false fired PostToolUseFailure with error="Exit code 1" — a non-zero Bash exit is routed to PostToolUseFailure UNLESS the harness re-interprets it — grep -c zzz /dev/null (exit 1) reached PostToolUse instead, with returnCodeInterpretation="No matches found" (Findings) (run tools)
• decision:block → honrada — hook_response: outcome=success exit_code=0; the reason PROBE_BLOCK reached the model/stream: yes (run posttool-block)
• systemMessage → honrada — marker PROBE_SYSMSG PostToolUse appears 2 times in the stream-json output (outside hook_response echoes: 5 informational;); in the transcript: 2 PROBE_SYSMSG PostToolBatch; 2 PROBE_SYSMSG PostToolUse; 2 PROBE_SYSMSG PreToolUse; 2 PROBE_SYSMSG SessionStart; 2 PROBE_SYSMSG Stop; 1 PROBE_SYSMSG UserPromptSubmit; (run system-message)
• additionalContext → honrada — the model repeated 'PROBE_CTX PostToolUse': no; the marker is in the transcript 3 time(s) (event fired in the run: 1 times) (run additional-context)

PostToolUseFailure — Claude Code

Disparou 1 vez(es) na(s) run(s) tools. hook_event_name: present; session_id: present. Chaves do payload: cwd, duration_ms, error, hook_event_name, is_interrupt, permission_mode, prompt_id, session_id, tool_input, tool_name, tool_use_id, transcript_path.

Payload, verbatim:

{"session_id":"777e68b4-1ad1-4f6c-9d81-64d0cae95a84","transcript_path":"/Users/jonhvmp/.claude/projects/-private-tmp-leopold-hookprobe-gjQnXe-claude-project/777e68b4-1ad1-4f6c-9d81-64d0cae95a84.jsonl","cwd":"/private/tmp/leopold-hookprobe.gjQnXe/claude-project","prompt_id":"4c67eb26-a782-4cfe-8201-a875dc0f7985","permission_mode":"bypassPermissions","hook_event_name":"PostToolUseFailure","tool_name":"Bash","tool_input":{"command":"false","description":"Step 2: Run false command"},"tool_use_id":"toolu_01QTAduNmuev93E45QX6sgsE","error":"Exit code 1","is_interrupt":false,"duration_ms":11}

Gatilhos tentados:

• run tools: a headless claude -p turn in a project with a CLAUDE.md that runs true, false, grep -c zzz /dev/null, an Edit of .leopold/PLAN.md and of the root PLAN.md, sleep 6 (the driver appends to both files from another process meanwhile), cat and, last, cd sub && pwd

PostToolBatch — Claude Code

Disparou 48 vez(es) na(s) run(s) additional-context, compact-auto, compact-manual-seed, config-change, config-change-exit2, filechanged-cwd, filechanged-nested, perm-allow, perm-auto, perm-deny, perm-host, posttool-block, pretool-deny, subagent, subagent-exit2, system-message, tasks, tasks-exit2, tools, worktree-agent. hook_event_name: present; session_id: present. Chaves do payload: agent_id, agent_type, cwd, hook_event_name, permission_mode, prompt_id, session_id, tool_calls, transcript_path.

Payload, verbatim:

{"session_id":"777e68b4-1ad1-4f6c-9d81-64d0cae95a84","transcript_path":"/Users/jonhvmp/.claude/projects/-private-tmp-leopold-hookprobe-gjQnXe-claude-project/777e68b4-1ad1-4f6c-9d81-64d0cae95a84.jsonl","cwd":"/private/tmp/leopold-hookprobe.gjQnXe/claude-project","prompt_id":"4c67eb26-a782-4cfe-8201-a875dc0f7985","permission_mode":"bypassPermissions","hook_event_name":"PostToolBatch","tool_calls":[{"tool_name":"Bash","tool_input":{"command":"true","description":"Step 1: Run true command"},"tool_use_id":"toolu_012pvg9iy5XR1HxLRHTMbFL8","tool_response":"(Bash completed with no output)"}]}

Gatilhos tentados:

• run tools: a headless claude -p turn in a project with a CLAUDE.md that runs true, false, grep -c zzz /dev/null, an Edit of .leopold/PLAN.md and of the root PLAN.md, sleep 6 (the driver appends to both files from another process meanwhile), cat and, last, cd sub && pwd

Respostas que o harness honrou:

• systemMessage → honrada — marker PROBE_SYSMSG PostToolBatch appears 2 times in the stream-json output (outside hook_response echoes: 5 informational;); in the transcript: 2 PROBE_SYSMSG PostToolBatch; 2 PROBE_SYSMSG PostToolUse; 2 PROBE_SYSMSG PreToolUse; 2 PROBE_SYSMSG SessionStart; 2 PROBE_SYSMSG Stop; 1 PROBE_SYSMSG UserPromptSubmit; (run system-message)
• additionalContext → honrada — the model repeated 'PROBE_CTX PostToolBatch': no; the marker is in the transcript 3 time(s) (event fired in the run: 1 times) (run additional-context)

Notification — Claude Code

Não disparou no Claude Code 2.1.259 (Claude Code). Gatilho tentado: run tools: a headless claude -p turn in a project with a CLAUDE.md that runs true, false, grep -c zzz /dev/null, an Edit of .leopold/PLAN.md and of the root PLAN.md, sleep 6 (the driver appends to both files from another process meanwhile), cat and, last, cd sub && pwd · runs perm-allow / perm-deny / perm-host: a permission decision in a headless session (matcher permission_prompt)

MessageDisplay — Claude Code

Disparou 48 vez(es) na(s) run(s) add-dir, additional-context, command, compact-auto, compact-manual-seed, config-change, config-change-exit2, fail-401, fail-429, fail-500, fail-529, filechanged-cwd, filechanged-nested, perm-allow, perm-auto, perm-deny, perm-host, posttool-block, pretool-deny, setup-init, setup-maintenance, stop-block, stop-exit2, subagent, subagent-exit2, system-message, tasks, tasks-exit2, teams, tools, unauth, worktree-agent. hook_event_name: present; session_id: present. Chaves do payload: cwd, delta, final, hook_event_name, index, message_id, prompt_id, session_id, transcript_path, turn_id.

Payload, verbatim:

{"session_id":"777e68b4-1ad1-4f6c-9d81-64d0cae95a84","transcript_path":"/Users/jonhvmp/.claude/projects/-private-tmp-leopold-hookprobe-gjQnXe-claude-project/777e68b4-1ad1-4f6c-9d81-64d0cae95a84.jsonl","cwd":"/private/tmp/leopold-hookprobe.gjQnXe/claude-project","prompt_id":"4c67eb26-a782-4cfe-8201-a875dc0f7985","hook_event_name":"MessageDisplay","turn_id":"73b5ff9c-0cac-4cde-9090-c8f3d5aa1339","message_id":"98f9e48b-9ef9-497e-9489-154e822c1cec","index":0,"final":true,"delta":"I'll execute these steps sequentially, one tool call per step."}

Gatilhos tentados:

• run tools: a headless claude -p turn in a project with a CLAUDE.md that runs true, false, grep -c zzz /dev/null, an Edit of .leopold/PLAN.md and of the root PLAN.md, sleep 6 (the driver appends to both files from another process meanwhile), cat and, last, cd sub && pwd

Respostas que o harness honrou:

• systemMessage → honrada — marker PROBE_SYSMSG MessageDisplay appears 1 times in the stream-json output (outside hook_response echoes: 5 informational;); in the transcript: 2 PROBE_SYSMSG PostToolBatch; 2 PROBE_SYSMSG PostToolUse; 2 PROBE_SYSMSG PreToolUse; 2 PROBE_SYSMSG SessionStart; 2 PROBE_SYSMSG Stop; 1 PROBE_SYSMSG UserPromptSubmit; (run system-message)

SubagentStart — Claude Code

Disparou 3 vez(es) na(s) run(s) subagent, subagent-exit2, worktree-agent. hook_event_name: present; session_id: present. Chaves do payload: agent_id, agent_type, cwd, hook_event_name, prompt_id, session_id, transcript_path.

Payload, verbatim:

{"session_id":"049615d1-d898-4fdb-ac17-53e8749b4b97","transcript_path":"/Users/jonhvmp/.claude/projects/-private-tmp-leopold-hookprobe-gjQnXe-claude-project/049615d1-d898-4fdb-ac17-53e8749b4b97.jsonl","cwd":"/private/tmp/leopold-hookprobe.gjQnXe/claude-project","prompt_id":"3ab1ba79-1a1c-47f8-8fd5-752a1894b69b","agent_id":"acb594f61d1631b05","agent_type":"general-purpose","hook_event_name":"SubagentStart"}

Gatilhos tentados:

• run subagent: the Agent tool (subagent_type general-purpose) running one Bash command

SubagentStop — Claude Code

Disparou 6 vez(es) na(s) run(s) compact-auto, compact-manual, subagent, subagent-exit2, worktree-agent. hook_event_name: present; session_id: present. Chaves do payload: agent_id, agent_transcript_path, agent_type, background_tasks, cwd, hook_event_name, last_assistant_message, permission_mode, prompt_id, session_crons, session_id, stop_hook_active, transcript_path.

Payload, verbatim (stop_hook_active: false):

{"session_id":"049615d1-d898-4fdb-ac17-53e8749b4b97","transcript_path":"/Users/jonhvmp/.claude/projects/-private-tmp-leopold-hookprobe-gjQnXe-claude-project/049615d1-d898-4fdb-ac17-53e8749b4b97.jsonl","cwd":"/private/tmp/leopold-hookprobe.gjQnXe/claude-project","prompt_id":"3ab1ba79-1a1c-47f8-8fd5-752a1894b69b","permission_mode":"bypassPermissions","agent_id":"acb594f61d1631b05","agent_type":"general-purpose","hook_event_name":"SubagentStop","stop_hook_active":false,"agent_transcript_path":"/Users/jonhvmp/.claude/projects/-private-tmp-leopold-hookprobe-gjQnXe-claude-project/049615d1-d898-4fdb-ac17-53e8749b4b97/subagents/agent-acb594f61d1631b05.jsonl","last_assistant_message":"PONG","background_tasks":[],"session_crons":[]}

Payload, verbatim (stop_hook_active: true):

{"session_id":"54b1292d-2d58-42bc-afaf-84baf89dc352","transcript_path":"/Users/jonhvmp/.claude/projects/-private-tmp-leopold-hookprobe-gjQnXe-claude-project/54b1292d-2d58-42bc-afaf-84baf89dc352.jsonl","cwd":"/private/tmp/leopold-hookprobe.gjQnXe/claude-project","prompt_id":"e57be662-12a6-46ed-85f2-8bf695bc7694","permission_mode":"bypassPermissions","agent_id":"af5a008acb95c5a66","agent_type":"general-purpose","hook_event_name":"SubagentStop","stop_hook_active":true,"agent_transcript_path":"/Users/jonhvmp/.claude/projects/-private-tmp-leopold-hookprobe-gjQnXe-claude-project/54b1292d-2d58-42bc-afaf-84baf89dc352/subagents/agent-af5a008acb95c5a66.jsonl","last_assistant_message":"Acknowledged. The subagent stop hook was executed successfully with:\n- Hook: `SubagentStop`\n- Output path: `/private/tmp/claude-501/-Users-jonhvmp-myspace-github-leopold/f8164f0a-f022-4498-9d87-6f196802cf56/scratchpad/final`\n- Result: `probe exit2 SubagentStop`\n\nIs there anything else you'd like me to do?","background_tasks":[],"session_crons":[]}

Gatilhos tentados:

• run subagent: the Agent tool (subagent_type general-purpose) running one Bash command

Respostas que o harness honrou:

• exit2 → honrada — SubagentStop fired 2 times in the run (exit 2 once, then the subagent stopped again); stop_hook_active on the second: true; hook_response: outcome=success exit_code=0 (run subagent-exit2)

TaskCreated — Claude Code

Disparou 2 vez(es) na(s) run(s) tasks, tasks-exit2. hook_event_name: present; session_id: present. Chaves do payload: cwd, hook_event_name, prompt_id, session_id, task_description, task_id, task_subject, transcript_path.

Payload, verbatim:

{"session_id":"24a348d1-5aa2-479f-aefe-51d20940d6cf","transcript_path":"/Users/jonhvmp/.claude/projects/-private-tmp-leopold-hookprobe-gjQnXe-claude-project/24a348d1-5aa2-479f-aefe-51d20940d6cf.jsonl","cwd":"/private/tmp/leopold-hookprobe.gjQnXe/claude-project","prompt_id":"f04c18bc-db45-4ccf-aebc-e0b6c15435b0","hook_event_name":"TaskCreated","task_id":"1","task_subject":"probe task","task_description":"probe"}

Gatilhos tentados:

• run tasks: TaskCreate then TaskUpdate(status=completed) then TaskList

TaskCompleted — Claude Code

Disparou 2 vez(es) na(s) run(s) tasks, tasks-exit2. hook_event_name: present; session_id: present. Chaves do payload: cwd, hook_event_name, prompt_id, session_id, task_description, task_id, task_subject, transcript_path.

Payload, verbatim:

{"session_id":"24a348d1-5aa2-479f-aefe-51d20940d6cf","transcript_path":"/Users/jonhvmp/.claude/projects/-private-tmp-leopold-hookprobe-gjQnXe-claude-project/24a348d1-5aa2-479f-aefe-51d20940d6cf.jsonl","cwd":"/private/tmp/leopold-hookprobe.gjQnXe/claude-project","prompt_id":"f04c18bc-db45-4ccf-aebc-e0b6c15435b0","hook_event_name":"TaskCompleted","task_id":"1","task_subject":"probe task","task_description":"probe"}

Gatilhos tentados:

• run tasks: TaskCreate then TaskUpdate(status=completed) then TaskList

Respostas que o harness honrou:

• exit2 → honrada — hook_response: outcome=error exit_code=2; final reply: Task created, updated to completed, and here's the status from TaskList: (run tasks-exit2)

Stop — Claude Code

Disparou 29 vez(es) na(s) run(s) add-dir, additional-context, command, compact-auto, compact-manual-seed, config-change, config-change-exit2, filechanged-cwd, filechanged-nested, perm-allow, perm-auto, perm-deny, perm-host, posttool-block, pretool-deny, setup-init, setup-maintenance, stop-block, stop-exit2, subagent, subagent-exit2, system-message, tasks, tasks-exit2, teams, tools, worktree-agent. hook_event_name: present; session_id: present. Chaves do payload: background_tasks, cwd, hook_event_name, last_assistant_message, permission_mode, prompt_id, session_crons, session_id, stop_hook_active, transcript_path.

Payload, verbatim (stop_hook_active: false):

{"session_id":"777e68b4-1ad1-4f6c-9d81-64d0cae95a84","transcript_path":"/Users/jonhvmp/.claude/projects/-private-tmp-leopold-hookprobe-gjQnXe-claude-project/777e68b4-1ad1-4f6c-9d81-64d0cae95a84.jsonl","cwd":"/private/tmp/leopold-hookprobe.gjQnXe/claude-project/sub","prompt_id":"4c67eb26-a782-4cfe-8201-a875dc0f7985","permission_mode":"bypassPermissions","hook_event_name":"Stop","stop_hook_active":false,"last_assistant_message":"DONE","background_tasks":[],"session_crons":[]}

Payload, verbatim (stop_hook_active: true):

{"session_id":"6e892317-009f-4b39-98c7-b8059db80916","transcript_path":"/Users/jonhvmp/.claude/projects/-private-tmp-leopold-hookprobe-gjQnXe-claude-project/6e892317-009f-4b39-98c7-b8059db80916.jsonl","cwd":"/private/tmp/leopold-hookprobe.gjQnXe/claude-project","prompt_id":"da942d3a-2805-441e-b5f2-17c165d0e5d8","permission_mode":"bypassPermissions","hook_event_name":"Stop","stop_hook_active":true,"last_assistant_message":"Understood. The stop hook has executed and captured the exit state. The probe shows `exit2 Stop` was recorded.\n\nWhat would you like me to do next? Are you testing the hook behavior, debugging the probe system, or is there a specific issue with the stop hook that needs investigation?","background_tasks":[],"session_crons":[]}

Gatilhos tentados:

• run tools: a headless claude -p turn in a project with a CLAUDE.md that runs true, false, grep -c zzz /dev/null, an Edit of .leopold/PLAN.md and of the root PLAN.md, sleep 6 (the driver appends to both files from another process meanwhile), cat and, last, cd sub && pwd

Respostas que o harness honrou:

• exit2 → honrada — Stop fired 2 times (exit 2 once; the model continued and stopped again); stop_hook_active on the second: true; hook_response: outcome=success exit_code=0 (run stop-exit2)
• decision:block → honrada — Stop fired 2 times ({"decision":"block"} once); stop_hook_active on the second: true (run stop-block)
• systemMessage → honrada — marker PROBE_SYSMSG Stop appears 2 times in the stream-json output (outside hook_response echoes: 5 informational;); in the transcript: 2 PROBE_SYSMSG PostToolBatch; 2 PROBE_SYSMSG PostToolUse; 2 PROBE_SYSMSG PreToolUse; 2 PROBE_SYSMSG SessionStart; 2 PROBE_SYSMSG Stop; 1 PROBE_SYSMSG UserPromptSubmit; (run system-message)

StopFailure — Claude Code

Disparou 5 vez(es) na(s) run(s) fail-500. hook_event_name: present; session_id: present. Chaves do payload: cwd, error, hook_event_name, last_assistant_message, prompt_id, session_id, transcript_path.

Payload, verbatim (error: "authentication_failed"):

{"session_id":"e175fdc4-0642-4929-adf4-f4d2b726fe67","transcript_path":"/Users/jonhvmp/.claude/projects/-private-tmp-leopold-hookprobe-gjQnXe-claude-project/e175fdc4-0642-4929-adf4-f4d2b726fe67.jsonl","cwd":"/private/tmp/leopold-hookprobe.gjQnXe/claude-project","prompt_id":"1403dcf9-1cdd-4834-a5f3-4e1fabdef016","hook_event_name":"StopFailure","error":"authentication_failed","last_assistant_message":"Failed to authenticate. API Error: 401 PROBE invalid credentials"}

Payload, verbatim (error: "rate_limit"):

{"session_id":"9a7d2b2e-5b6d-4db8-b454-40ddd4c47b40","transcript_path":"/Users/jonhvmp/.claude/projects/-private-tmp-leopold-hookprobe-gjQnXe-claude-project/9a7d2b2e-5b6d-4db8-b454-40ddd4c47b40.jsonl","cwd":"/private/tmp/leopold-hookprobe.gjQnXe/claude-project","prompt_id":"4298ce5d-a63c-4816-890d-1eeea7d5d0df","hook_event_name":"StopFailure","error":"rate_limit","last_assistant_message":"API Error: Server is temporarily limiting requests (not your usage limit) · PROBE rate limited"}

Payload, verbatim (error: "server_error"):

{"session_id":"5189f5c3-682d-4a38-af63-21fa311293bb","transcript_path":"/Users/jonhvmp/.claude/projects/-private-tmp-leopold-hookprobe-gjQnXe-claude-project/5189f5c3-682d-4a38-af63-21fa311293bb.jsonl","cwd":"/private/tmp/leopold-hookprobe.gjQnXe/claude-project","prompt_id":"20cb75b2-493f-4666-817a-84c47154ca60","hook_event_name":"StopFailure","error":"server_error","last_assistant_message":"API Error: 500 PROBE internal server error. This is a server-side issue, usually temporary — try again in a moment. If it persists, check your inference gateway (127.0.0.1:19200)."}

Gatilhos tentados:

• run fail-401: ANTHROPIC_BASE_URL at a Python-stdlib stub answering HTTP 401 (CLAUDE_CODE_MAX_RETRIES=0)
• run fail-429: ANTHROPIC_BASE_URL at a Python-stdlib stub answering HTTP 429 (CLAUDE_CODE_MAX_RETRIES=0)
• run fail-500: ANTHROPIC_BASE_URL at a Python-stdlib stub answering HTTP 500 (CLAUDE_CODE_MAX_RETRIES=0)
• run fail-529: ANTHROPIC_BASE_URL at a Python-stdlib stub answering HTTP 529 (CLAUDE_CODE_MAX_RETRIES=0)
• run unauth: a fresh CLAUDE_CONFIG_DIR with no login (Not logged in)

Respostas que o harness honrou:

• observe → n/a — HTTP 401 from the stub (1 request(s) to /v1/messages?beta=true) → StopFailure error=; Stop fired: 0; SessionEnd fired: 0 (run fail-401)
• observe → n/a — HTTP 429 from the stub (1 request(s) to /v1/messages?beta=true) → StopFailure error=; Stop fired: 0; SessionEnd fired: 1 (run fail-429)
• observe → n/a — HTTP 500 from the stub (1 request(s) to /v1/messages?beta=true) → StopFailure error=authentication_failed; Stop fired: 0; SessionEnd fired: 0 (run fail-500)
• observe → n/a — HTTP 529 from the stub (1 request(s) to /v1/messages?beta=true) → StopFailure error=; Stop fired: 0; SessionEnd fired: 0 (run fail-529)
• observe → n/a — unauthenticated temp config dir: StopFailure error=; events that fired: InstructionsLoaded MessageDisplay SessionStart UserPromptSubmit (run unauth)

TeammateIdle — Claude Code

Não disparou no Claude Code 2.1.259 (Claude Code). Gatilho tentado: run teams: CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS=1 and a prompt asking for TeamCreate + one teammate going idle (reply: NO_TEAM_TOOL)

InstructionsLoaded — Claude Code

Disparou 38 vez(es) na(s) run(s) add-dir, additional-context, command, command-exit2, compact-auto, compact-manual, compact-manual-seed, config-change, config-change-exit2, fail-401, fail-429, fail-500, fail-529, filechanged-cwd, filechanged-nested, model-switch, model-switch-exit2, perm-allow, perm-auto, perm-deny, perm-host, posttool-block, pretool-deny, prompt-exit2, setup-init, setup-maintenance, stop-block, stop-exit2, subagent, subagent-exit2, system-message, tasks, tasks-exit2, teams, tools, unauth, worktree-agent. hook_event_name: present; session_id: present. Chaves do payload: cwd, file_path, hook_event_name, load_reason, memory_type, session_id, transcript_path.

Payload, verbatim:

{"session_id":"777e68b4-1ad1-4f6c-9d81-64d0cae95a84","transcript_path":"/Users/jonhvmp/.claude/projects/-private-tmp-leopold-hookprobe-gjQnXe-claude-project/777e68b4-1ad1-4f6c-9d81-64d0cae95a84.jsonl","cwd":"/private/tmp/leopold-hookprobe.gjQnXe/claude-project","hook_event_name":"InstructionsLoaded","file_path":"/private/tmp/leopold-hookprobe.gjQnXe/claude-project/CLAUDE.md","memory_type":"Project","load_reason":"session_start"}

Gatilhos tentados:

• run tools: a headless claude -p turn in a project with a CLAUDE.md that runs true, false, grep -c zzz /dev/null, an Edit of .leopold/PLAN.md and of the root PLAN.md, sleep 6 (the driver appends to both files from another process meanwhile), cat and, last, cd sub && pwd

Respostas que o harness honrou:

• systemMessage → não observável — marker PROBE_SYSMSG InstructionsLoaded appears 0 times in the stream-json output (outside hook_response echoes: 5 informational;); in the transcript: 2 PROBE_SYSMSG PostToolBatch; 2 PROBE_SYSMSG PostToolUse; 2 PROBE_SYSMSG PreToolUse; 2 PROBE_SYSMSG SessionStart; 2 PROBE_SYSMSG Stop; 1 PROBE_SYSMSG UserPromptSubmit; (run system-message)
• additionalContext → não honrada — the model repeated 'PROBE_CTX InstructionsLoaded': no; the marker is in the transcript 0 0 time(s) (event fired in the run: 1 times) (run additional-context)

ConfigChange — Claude Code

Disparou 2 vez(es) na(s) run(s) config-change, config-change-exit2. hook_event_name: present; session_id: present. Chaves do payload: cwd, file_path, hook_event_name, prompt_id, session_id, source, transcript_path.

Payload, verbatim:

{"session_id":"23d889d0-a993-443d-8c40-260d3d2d2180","transcript_path":"/Users/jonhvmp/.claude/projects/-private-tmp-leopold-hookprobe-gjQnXe-claude-project/23d889d0-a993-443d-8c40-260d3d2d2180.jsonl","cwd":"/private/tmp/leopold-hookprobe.gjQnXe/claude-project","prompt_id":"2df4e6ad-452e-4eda-a97c-4e395086d0a9","hook_event_name":"ConfigChange","source":"project_settings","file_path":"/private/tmp/leopold-hookprobe.gjQnXe/claude-project/.claude/settings.json"}

Gatilhos tentados:

• run config-change: the session's own Bash rewrites .claude/settings.json mid-session (jq adds a key), then sleeps 3s

Respostas que o harness honrou:

• exit2 → não observável — hook_response: ; the file on disk keeps the edit (the block stops the reload, not the write): probeKey present = true (run config-change-exit2)

CwdChanged — Claude Code

Disparou 2 vez(es) na(s) run(s) filechanged-cwd, tools. hook_event_name: present; session_id: present. Chaves do payload: cwd, hook_event_name, new_cwd, old_cwd, prompt_id, session_id, transcript_path.

Payload, verbatim:

{"session_id":"777e68b4-1ad1-4f6c-9d81-64d0cae95a84","transcript_path":"/Users/jonhvmp/.claude/projects/-private-tmp-leopold-hookprobe-gjQnXe-claude-project/777e68b4-1ad1-4f6c-9d81-64d0cae95a84.jsonl","cwd":"/private/tmp/leopold-hookprobe.gjQnXe/claude-project/sub","prompt_id":"4c67eb26-a782-4cfe-8201-a875dc0f7985","hook_event_name":"CwdChanged","old_cwd":"/private/tmp/leopold-hookprobe.gjQnXe/claude-project","new_cwd":"/private/tmp/leopold-hookprobe.gjQnXe/claude-project/sub"}

Gatilhos tentados:

• run tools: a headless claude -p turn in a project with a CLAUDE.md that runs true, false, grep -c zzz /dev/null, an Edit of .leopold/PLAN.md and of the root PLAN.md, sleep 6 (the driver appends to both files from another process meanwhile), cat and, last, cd sub && pwd

Respostas que o harness honrou:

• systemMessage → não observável — marker PROBE_SYSMSG CwdChanged appears 0 times in the stream-json output (outside hook_response echoes: 5 informational;); in the transcript: 2 PROBE_SYSMSG PostToolBatch; 2 PROBE_SYSMSG PostToolUse; 2 PROBE_SYSMSG PreToolUse; 2 PROBE_SYSMSG SessionStart; 2 PROBE_SYSMSG Stop; 1 PROBE_SYSMSG UserPromptSubmit; (run system-message)

DirectoryAdded — Claude Code

Não disparou no Claude Code 2.1.259 (Claude Code). Gatilho tentado: run add-dir: the /add-dir command in -p (answer: /add-dir isn't available in this environment.); run add-dir-control: an SDK control_request register_repo_root over stream-json stdin (response: "error")

FileChanged — Claude Code

Disparou 6 vez(es) na(s) run(s) filechanged-nested, tools. hook_event_name: present; session_id: present. Chaves do payload: cwd, event, file_path, hook_event_name, prompt_id, session_id, transcript_path.

Payload, verbatim (file_path: "/private/tmp/leopold-hookprobe.gjQnXe/claude-project/.leopold/PLAN.md"):

{"session_id":"777e68b4-1ad1-4f6c-9d81-64d0cae95a84","transcript_path":"/Users/jonhvmp/.claude/projects/-private-tmp-leopold-hookprobe-gjQnXe-claude-project/777e68b4-1ad1-4f6c-9d81-64d0cae95a84.jsonl","cwd":"/private/tmp/leopold-hookprobe.gjQnXe/claude-project","prompt_id":"4c67eb26-a782-4cfe-8201-a875dc0f7985","hook_event_name":"FileChanged","file_path":"/private/tmp/leopold-hookprobe.gjQnXe/claude-project/.leopold/PLAN.md","event":"change"}

Payload, verbatim (file_path: "/private/tmp/leopold-hookprobe.gjQnXe/claude-project/PLAN.md"):

{"session_id":"777e68b4-1ad1-4f6c-9d81-64d0cae95a84","transcript_path":"/Users/jonhvmp/.claude/projects/-private-tmp-leopold-hookprobe-gjQnXe-claude-project/777e68b4-1ad1-4f6c-9d81-64d0cae95a84.jsonl","cwd":"/private/tmp/leopold-hookprobe.gjQnXe/claude-project","prompt_id":"4c67eb26-a782-4cfe-8201-a875dc0f7985","hook_event_name":"FileChanged","file_path":"/private/tmp/leopold-hookprobe.gjQnXe/claude-project/PLAN.md","event":"change"}

Gatilhos tentados:

• run filechanged-cwd: the same wiring, with cd sub (CwdChanged) BEFORE the edits and the external append
• run filechanged-nested: the same two entries with NO PLAN.md in the project root — an Edit of .leopold/PLAN.md plus an append from another process
• run tools: a headless claude -p turn in a project with a CLAUDE.md that runs true, false, grep -c zzz /dev/null, an Edit of .leopold/PLAN.md and of the root PLAN.md, sleep 6 (the driver appends to both files from another process meanwhile), cat and, last, cd sub && pwd

Respostas que o harness honrou:

• observe → n/a — two entries wired: matcher .leopold/PLAN.md (registers the watch) and matcher PLAN.md (receives); firings per entry and path: 2 matches=PLAN.md -> .leopold/PLAN.md; 2 matches=PLAN.md -> PLAN.md; — the session's own Edit and the other process's append produce payloads of the same shape (file_path + event), nothing in the payload tells them apart; the path-shaped entry never fires itself (run tools)
• observe (after a cd in the session) → n/a — after cd sub (CwdChanged) the same wiring fired 0 times for two Edits and one external append — a cwd change detaches the watches (run filechanged-cwd)
• observe (no root PLAN.md) → n/a — with no PLAN.md in the project root, FileChanged fired 2 times for .leopold/PLAN.md (Edit + external append): 2 .leopold/PLAN.md; (run filechanged-nested)

WorktreeCreate — Claude Code

Disparou 2 vez(es) na(s) run(s) worktree, worktree-agent. hook_event_name: present; session_id: present. Chaves do payload: cwd, hook_event_name, name, prompt_id, session_id, transcript_path.

Payload, verbatim:

{"session_id":"cb14dfc0-a66d-486e-aeec-c0d6d37375c3","transcript_path":"/Users/jonhvmp/.claude/projects/-private-tmp-leopold-hookprobe-gjQnXe-claude-project/cb14dfc0-a66d-486e-aeec-c0d6d37375c3.jsonl","cwd":"/private/tmp/leopold-hookprobe.gjQnXe/claude-project","hook_event_name":"WorktreeCreate","name":"probe-wt"}

Gatilhos tentados:

• run worktree: claude -p --worktree probe-wt in a git project (the hook creates the worktree and echoes its path — an empty answer aborts the session)

Respostas que o harness honrou:

• worktree → honrada — the session ran inside the path the hook echoed: yes; worktrees after the run: 2 (run worktree)

WorktreeRemove — Claude Code

Não disparou no Claude Code 2.1.259 (Claude Code). Gatilho tentado: run worktree: the --worktree session exiting with a clean worktree; run worktree-agent: an Agent tool call with isolation worktree finishing

PreCompact — Claude Code

Disparou 2 vez(es) na(s) run(s) compact-auto, compact-manual. hook_event_name: present; session_id: present. Chaves do payload: custom_instructions, cwd, hook_event_name, prompt_id, session_id, transcript_path, trigger.

Payload, verbatim (trigger: "auto"):

{"session_id":"8696992c-e252-4b26-aa6e-0f09b973a8fc","transcript_path":"/Users/jonhvmp/.claude/projects/-private-tmp-leopold-hookprobe-gjQnXe-claude-project/8696992c-e252-4b26-aa6e-0f09b973a8fc.jsonl","cwd":"/private/tmp/leopold-hookprobe.gjQnXe/claude-project","prompt_id":"b1988ab3-5c85-48da-81fd-b0bd3a241365","hook_event_name":"PreCompact","trigger":"auto","custom_instructions":null}

Payload, verbatim (trigger: "manual"):

{"session_id":"ea4f4119-202c-4931-9721-72c83db1f7af","transcript_path":"/Users/jonhvmp/.claude/projects/-private-tmp-leopold-hookprobe-gjQnXe-claude-project/ea4f4119-202c-4931-9721-72c83db1f7af.jsonl","cwd":"/private/tmp/leopold-hookprobe.gjQnXe/claude-project","prompt_id":"5dc35203-ab1c-4bb5-900f-f1a67c33d899","hook_event_name":"PreCompact","trigger":"manual","custom_instructions":null}

Gatilhos tentados:

• run compact-auto: --autocompact 100000 and three full reads of a 120 KB file (CLAUDE_CODE_AUTO_COMPACT_WINDOW=60000 + CLAUDE_AUTOCOMPACT_PCT_OVERRIDE=50 fired both events too in the pilot)
• run compact-manual: claude -p --resume <id> /compact

Respostas que o harness honrou:

• observe → n/a — auto compaction: PreCompact fired 1 times, PostCompact 1 times; compact_boundary in the stream: {"trigger":"auto","pre_tokens":70493,"post_tokens":7859}; SessionStart fired again after each compaction (source: compact,startup,) (run compact-auto)
• observe → n/a — manual compaction via /compact on a resumed session: PreCompact 1, PostCompact 1 (trigger: manual) (run compact-manual)

PostCompact — Claude Code

Disparou 2 vez(es) na(s) run(s) compact-auto, compact-manual. hook_event_name: present; session_id: present. Chaves do payload: compact_summary, cwd, hook_event_name, prompt_id, session_id, transcript_path, trigger.

Payload, verbatim (trigger: "auto"):

{"session_id":"8696992c-e252-4b26-aa6e-0f09b973a8fc","transcript_path":"/Users/jonhvmp/.claude/projects/-private-tmp-leopold-hookprobe-gjQnXe-claude-project/8696992c-e252-4b26-aa6e-0f09b973a8fc.jsonl","cwd":"/private/tmp/leopold-hookprobe.gjQnXe/claude-project","prompt_id":"b1988ab3-5c85-48da-81fd-b0bd3a241365","hook_event_name":"PostCompact","trigger":"auto","compact_summary":"<analysis>\nLet me carefully analyze this conversation chronologically:\n\n1. **Initial Setup**: The conversation begins with system context showing a git repository with branch \"main\" and some modified/untracked files (.leopold/PLAN.md, PLAN.md, .claude/settings.json files, probe-allow.txt).\n\n2. **User's Primary Request**: The user explicitly requests:\n   - Read big.txt (whole file, no offset/limit) three separate times, one call at a time\n   - Run Bash command: wc -c big.txt\n …[+4237 chars]"}

Payload, verbatim (trigger: "manual"):

{"session_id":"ea4f4119-202c-4931-9721-72c83db1f7af","transcript_path":"/Users/jonhvmp/.claude/projects/-private-tmp-leopold-hookprobe-gjQnXe-claude-project/ea4f4119-202c-4931-9721-72c83db1f7af.jsonl","cwd":"/private/tmp/leopold-hookprobe.gjQnXe/claude-project","prompt_id":"5dc35203-ab1c-4bb5-900f-f1a67c33d899","hook_event_name":"PostCompact","trigger":"manual","compact_summary":"<analysis>\nThe conversation history is minimal and focused on two distinct requests:\n\n1. Initial Request (Message 1): The user asked me to read big.txt in its entirety and respond with \"DONE\".\n   - I executed this request using the Read tool\n   - The Read tool returned a partial view due to character limits (showing first 24527 of 120001 characters)\n   - The file contains what appears to be encoded or random string data\n\n2. Second Request (Current Message): The user is now as…[+3716 chars]"}

Gatilhos tentados:

• run compact-auto: --autocompact 100000 and three full reads of a 120 KB file (CLAUDE_CODE_AUTO_COMPACT_WINDOW=60000 + CLAUDE_AUTOCOMPACT_PCT_OVERRIDE=50 fired both events too in the pilot)
• run compact-manual: claude -p --resume <id> /compact

PreModelSwitch — Claude Code

Disparou 2 vez(es) na(s) run(s) model-switch, model-switch-exit2. hook_event_name: present; session_id: present. Chaves do payload: cache_ttl, context_tokens, cwd, estimated_cache_write_usd, from_model, hook_event_name, pricing, prompt_cache_warm, prompt_id, requested_model, session_id, source, to_model, transcript_path.

Payload, verbatim:

{"session_id":"1d5b76b2-f059-4f62-8a9d-48a0f9f6f1c9","transcript_path":"/Users/jonhvmp/.claude/projects/-private-tmp-leopold-hookprobe-gjQnXe-claude-project/1d5b76b2-f059-4f62-8a9d-48a0f9f6f1c9.jsonl","cwd":"/private/tmp/leopold-hookprobe.gjQnXe/claude-project","prompt_id":"3112dde5-323d-492f-aaf9-4843adb1d9af","hook_event_name":"PreModelSwitch","from_model":"claude-haiku-4-5-20251001","to_model":"claude-sonnet-5","requested_model":"sonnet","source":"command","context_tokens":0,"prompt_cache_warm":false,"cache_ttl":"1h","estimated_cache_write_usd":0,"pricing":"catalog"}

Gatilhos tentados:

• run model-switch: the /model sonnet command in a -p session started on haiku

Respostas que o harness honrou:

• exit2 → honrada — PostModelSwitch fired 0 times after the block (0 = the switch did not happen); final: Model switch to Sonnet 5 was blocked by a PreModelSwitch hook: [bash /Users/jonhvmp/myspace/github/leopold/scripts/probe/dump-hook.sh PreModelSwitch --out /priv (run model-switch-exit2)

PostModelSwitch — Claude Code

Disparou 1 vez(es) na(s) run(s) model-switch. hook_event_name: present; session_id: present. Chaves do payload: cache_ttl, context_tokens, cwd, estimated_cache_write_usd, from_model, hook_event_name, pricing, prompt_cache_warm, prompt_id, requested_model, session_id, source, to_model, transcript_path.

Payload, verbatim:

{"session_id":"1d5b76b2-f059-4f62-8a9d-48a0f9f6f1c9","transcript_path":"/Users/jonhvmp/.claude/projects/-private-tmp-leopold-hookprobe-gjQnXe-claude-project/1d5b76b2-f059-4f62-8a9d-48a0f9f6f1c9.jsonl","cwd":"/private/tmp/leopold-hookprobe.gjQnXe/claude-project","prompt_id":"3112dde5-323d-492f-aaf9-4843adb1d9af","hook_event_name":"PostModelSwitch","from_model":"claude-haiku-4-5-20251001","to_model":"claude-sonnet-5","requested_model":"sonnet","source":"command","context_tokens":0,"prompt_cache_warm":false,"cache_ttl":"1h","estimated_cache_write_usd":0,"pricing":"catalog"}

Gatilhos tentados:

• run model-switch: the /model sonnet command in a -p session started on haiku

Elicitation — Claude Code

Não disparou no Claude Code 2.1.259 (Claude Code). Gatilho tentado: run elicit-observe: a stdio MCP stub whose tool sends elicitation/create back to the client

Respostas que o harness honrou:

• action:accept+content → não honrada — the MCP server received the hook's answer: ; ElicitationResult fired: 0 (it fires only on the user/cancel path) (run elicit-accept)

ElicitationResult — Claude Code

Não disparou no Claude Code 2.1.259 (Claude Code). Gatilho tentado: run elicit-observe: the same elicitation, unanswered by any hook, in a headless session (auto-cancelled)

Respostas que o harness honrou:

• observe → n/a — with no hook answer the headless session cancelled: action=; the MCP server received: (run elicit-observe)

SessionEnd — Claude Code

Disparou 34 vez(es) na(s) run(s) add-dir, additional-context, command, command-exit2, compact-auto, compact-manual, compact-manual-seed, config-change, config-change-exit2, fail-429, filechanged-cwd, filechanged-nested, model-switch, model-switch-exit2, perm-allow, perm-auto, perm-deny, perm-host, posttool-block, pretool-deny, prompt-exit2, setup-init, setup-maintenance, stop-block, stop-exit2, subagent, subagent-exit2, system-message, tasks, tasks-exit2, teams, tools, worktree-agent. hook_event_name: present; session_id: present. Chaves do payload: cwd, hook_event_name, prompt_id, reason, session_id, transcript_path.

Payload, verbatim:

{"session_id":"777e68b4-1ad1-4f6c-9d81-64d0cae95a84","transcript_path":"/Users/jonhvmp/.claude/projects/-private-tmp-leopold-hookprobe-gjQnXe-claude-project/777e68b4-1ad1-4f6c-9d81-64d0cae95a84.jsonl","cwd":"/private/tmp/leopold-hookprobe.gjQnXe/claude-project/sub","prompt_id":"4c67eb26-a782-4cfe-8201-a875dc0f7985","hook_event_name":"SessionEnd","reason":"other"}

Gatilhos tentados:

• run tools: a headless claude -p turn in a project with a CLAUDE.md that runs true, false, grep -c zzz /dev/null, an Edit of .leopold/PLAN.md and of the root PLAN.md, sleep 6 (the driver appends to both files from another process meanwhile), cat and, last, cd sub && pwd

Respostas que o harness honrou:

• systemMessage → não observável — marker PROBE_SYSMSG SessionEnd appears 0 times in the stream-json output (outside hook_response echoes: 5 informational;); in the transcript: 2 PROBE_SYSMSG PostToolBatch; 2 PROBE_SYSMSG PostToolUse; 2 PROBE_SYSMSG PreToolUse; 2 PROBE_SYSMSG SessionStart; 2 PROBE_SYSMSG Stop; 1 PROBE_SYSMSG UserPromptSubmit; (run system-message)

Eventos — Codex CLI

SessionStart — Codex CLI

Disparou 25 vez(es) na(s) run(s) additional-context, compact, exec-as-role-4, fail-401, fail-429, fail-500, fail-529, interrupt, perm-allow, perm-deny, perm-onrequest, posttool-block, pretool-deny, prompt-exit2, role-bogus, role-bogus-strict, stop-block, stop-exit2, strict-config, subagent, subagent-exit2, subagent-role, system-message, tools. hook_event_name: present; session_id: present. Chaves do payload: cwd, hook_event_name, model, permission_mode, session_id, source, transcript_path.

Payload, verbatim (source: "startup"):

{"session_id":"01a064ee-a7c6-7af3-872f-dd265be1acc2","transcript_path":"/private/tmp/leopold-hookprobe.gjQnXe/codex-home/sessions/2026/09/02/rollout-2026-09-02T22-42-37-01a064ee-a7c6-7af3-872f-dd265be1acc2.jsonl","cwd":"/tmp/leopold-hookprobe.gjQnXe/codex-project","hook_event_name":"SessionStart","model":"gpt-5.6-sol","permission_mode":"bypassPermissions","source":"startup"}

Payload, verbatim (source: "compact"):

{"session_id":"01a064f2-3530-7453-9e93-f1a7093bb39d","transcript_path":"/private/tmp/leopold-hookprobe.gjQnXe/codex-home/sessions/2026/09/02/rollout-2026-09-02T22-46-30-01a064f2-3530-7453-9e93-f1a7093bb39d.jsonl","cwd":"/tmp/leopold-hookprobe.gjQnXe/codex-project","hook_event_name":"SessionStart","model":"gpt-5.6-sol","permission_mode":"bypassPermissions","source":"compact"}

Gatilhos tentados:

• run tools: a headless codex exec turn in a project with an AGENTS.md that runs true, false, grep -c zzz /dev/null, an apply_patch edit of .leopold/PLAN.md, sleep 6 (the driver appends to .leopold/PLAN.md from another process meanwhile) and cat

Respostas que o harness honrou:

• systemMessage → não observável — marker PROBE_SYSMSG SessionStart in the --json stream: 0 times; in the rollout transcript: 0 0 times (run system-message)
• additionalContext → não honrada — the model repeated 'PROBE_CTX SessionStart': no (run additional-context)
• --strict-config with the probe's hook block → n/a — rc=0; the hook wiring the shared writer emits passes --strict-config: yes (run strict-config)

SessionEnd — Codex CLI

Disparou 24 vez(es) na(s) run(s) additional-context, compact, exec-as-role-4, fail-401, fail-429, fail-500, fail-529, interrupt, perm-allow, perm-deny, perm-onrequest, posttool-block, pretool-deny, prompt-exit2, role-bogus, role-bogus-strict, stop-block, stop-exit2, strict-config, subagent, subagent-exit2, subagent-role, system-message, tools. hook_event_name: present; session_id: present. Chaves do payload: cwd, hook_event_name, reason, session_id, transcript_path.

Payload, verbatim:

{"session_id":"01a064ee-a7c6-7af3-872f-dd265be1acc2","transcript_path":"/private/tmp/leopold-hookprobe.gjQnXe/codex-home/sessions/2026/09/02/rollout-2026-09-02T22-42-37-01a064ee-a7c6-7af3-872f-dd265be1acc2.jsonl","cwd":"/tmp/leopold-hookprobe.gjQnXe/codex-project","hook_event_name":"SessionEnd","reason":"other"}

Gatilhos tentados:

• run tools: a headless codex exec turn in a project with an AGENTS.md that runs true, false, grep -c zzz /dev/null, an apply_patch edit of .leopold/PLAN.md, sleep 6 (the driver appends to .leopold/PLAN.md from another process meanwhile) and cat

Respostas que o harness honrou:

• systemMessage → não observável — marker PROBE_SYSMSG SessionEnd in the --json stream: 0 times; in the rollout transcript: 0 0 times (run system-message)

SubagentStart — Codex CLI

Disparou 3 vez(es) na(s) run(s) subagent, subagent-exit2, subagent-role. hook_event_name: present; session_id: present. Chaves do payload: agent_id, agent_type, cwd, hook_event_name, model, permission_mode, session_id, transcript_path, turn_id.

Payload, verbatim (agent_type: "default"):

{"session_id":"01a064ef-43f8-7a31-a2c4-e372c36fb4bd","turn_id":"01a064ef-557b-76b3-b60f-c3e732663d6c","transcript_path":"/private/tmp/leopold-hookprobe.gjQnXe/codex-home/sessions/2026/09/02/rollout-2026-09-02T22-43-21-01a064ef-556c-7670-a0f2-9f6a2fe2c42a.jsonl","cwd":"/tmp/leopold-hookprobe.gjQnXe/codex-project","hook_event_name":"SubagentStart","model":"gpt-5.6-sol","permission_mode":"bypassPermissions","agent_id":"01a064ef-556c-7670-a0f2-9f6a2fe2c42a","agent_type":"default"}

Payload, verbatim (agent_type: "probe-reviewer"):

{"session_id":"01a064ef-8fc9-7782-b75b-cfdcb7ee8a12","turn_id":"01a064ef-af79-77d2-99ca-62914c6c39a4","transcript_path":"/private/tmp/leopold-hookprobe.gjQnXe/codex-home/sessions/2026/09/02/rollout-2026-09-02T22-43-44-01a064ef-af6b-73a1-b2d6-9d4f09708e5d.jsonl","cwd":"/tmp/leopold-hookprobe.gjQnXe/codex-project","hook_event_name":"SubagentStart","model":"gpt-5.6-sol","permission_mode":"bypassPermissions","agent_id":"01a064ef-af6b-73a1-b2d6-9d4f09708e5d","agent_type":"probe-reviewer"}

Gatilhos tentados:

• run subagent-role: spawn_agent with agent_type probe-reviewer (.codex/agents/probe-reviewer.toml)
• run subagent: spawn_agent (default role) + wait_agent

Respostas que o harness honrou:

• observe → n/a — spawn tool name as seen by PreToolUse: collaborationspawn_agent; SubagentStop carries agent_transcript_path: true (run subagent)
• role file → honrada — agent_type in SubagentStart: probe-reviewer; the role's fixed reply came back: yes (run subagent-role)
• role file with an unknown key → n/a — without --strict-config: Ignoring malformed agent role definition: failed to deserialize agent role file at /tmp/leopold-hookprobe.gjQnXe/codex-project/.codex/agents/bogus.toml: unknown field bogus_key ; reply: unknown agent_type 'bogus' (run role-bogus)
• role file with an unknown key (--strict-config) → n/a — with --strict-config: Ignoring malformed agent role definition: failed to deserialize agent role file at /tmp/leopold-hookprobe.gjQnXe/codex-project/.codex/agents/bogus.toml: unknown field bogus_key ; reply: unknown agent_type 'bogus' (run role-bogus-strict)
• codex exec -c agent_role="probe-reviewer" → n/a — rc=1; reply: ; stderr: Error loading config.toml: unknown configuration field agent_role in -c/--config override (run exec-as-role-1)
• codex exec -c agent_type="probe-reviewer" → n/a — rc=1; reply: ; stderr: Error loading config.toml: unknown configuration field agent_type in -c/--config override (run exec-as-role-2)
• codex exec -c role="probe-reviewer" → n/a — rc=1; reply: ; stderr: Error loading config.toml: unknown configuration field role in -c/--config override (run exec-as-role-3)
• codex exec -c agents.probe-reviewer.config_file=".codex/agents/probe-reviewer.toml" → n/a — rc=0; reply: NO_ROLE; stderr: (run exec-as-role-4)
• codex exec -p <role file as profile> → n/a — rc=1; stderr: Error loading config.toml: /private/tmp/leopold-hookprobe.gjQnXe/codex-home/probe-reviewer.config.toml:1:1: unknown configuration field name (run exec-as-role-profile)

SubagentStop — Codex CLI

Disparou 4 vez(es) na(s) run(s) subagent, subagent-exit2, subagent-role. hook_event_name: present; session_id: present. Chaves do payload: agent_id, agent_transcript_path, agent_type, cwd, hook_event_name, last_assistant_message, model, permission_mode, session_id, stop_hook_active, transcript_path, turn_id.

Payload, verbatim (stop_hook_active: false):

{"session_id":"01a064ef-43f8-7a31-a2c4-e372c36fb4bd","turn_id":"01a064ef-557b-76b3-b60f-c3e732663d6c","transcript_path":"/private/tmp/leopold-hookprobe.gjQnXe/codex-home/sessions/2026/09/02/rollout-2026-09-02T22-43-17-01a064ef-43f8-7a31-a2c4-e372c36fb4bd.jsonl","agent_transcript_path":"/private/tmp/leopold-hookprobe.gjQnXe/codex-home/sessions/2026/09/02/rollout-2026-09-02T22-43-21-01a064ef-556c-7670-a0f2-9f6a2fe2c42a.jsonl","cwd":"/tmp/leopold-hookprobe.gjQnXe/codex-project","hook_event_name":"SubagentStop","model":"gpt-5.6-sol","permission_mode":"bypassPermissions","stop_hook_active":false,"agent_id":"01a064ef-556c-7670-a0f2-9f6a2fe2c42a","agent_type":"default","last_assistant_message":"PONG"}

Payload, verbatim (stop_hook_active: true):

{"session_id":"01a064f0-23a6-7082-a96c-4b3b8175e81b","turn_id":"01a064f0-35b9-7352-bbc3-b7de53c28f04","transcript_path":"/private/tmp/leopold-hookprobe.gjQnXe/codex-home/sessions/2026/09/02/rollout-2026-09-02T22-44-14-01a064f0-23a6-7082-a96c-4b3b8175e81b.jsonl","agent_transcript_path":"/private/tmp/leopold-hookprobe.gjQnXe/codex-home/sessions/2026/09/02/rollout-2026-09-02T22-44-19-01a064f0-35ac-77d1-9871-fd30fa81b96c.jsonl","cwd":"/tmp/leopold-hookprobe.gjQnXe/codex-project","hook_event_name":"SubagentStop","model":"gpt-5.6-sol","permission_mode":"bypassPermissions","stop_hook_active":true,"agent_id":"01a064f0-35ac-77d1-9871-fd30fa81b96c","agent_type":"default","last_assistant_message":"PONG"}

Gatilhos tentados:

• run subagent: spawn_agent (default role) + wait_agent

Respostas que o harness honrou:

• exit2 → honrada — SubagentStop fired 2 times (exit 2 once); stop_hook_active on the second: true (run subagent-exit2)

PreToolUse — Codex CLI

Disparou 29 vez(es) na(s) run(s) additional-context, compact, interrupt, perm-allow, perm-deny, perm-onrequest, posttool-block, pretool-deny, role-bogus, role-bogus-strict, subagent, subagent-exit2, subagent-role, system-message, tools. hook_event_name: present; session_id: present. Chaves do payload: cwd, hook_event_name, model, permission_mode, session_id, tool_input, tool_name, tool_use_id, transcript_path, turn_id.

Payload, verbatim (Bash: {"command":"true"}):

{"session_id":"01a064ee-a7c6-7af3-872f-dd265be1acc2","turn_id":"01a064ee-a831-76d0-bfad-138dcf2efcf2","transcript_path":"/private/tmp/leopold-hookprobe.gjQnXe/codex-home/sessions/2026/09/02/rollout-2026-09-02T22-42-37-01a064ee-a7c6-7af3-872f-dd265be1acc2.jsonl","cwd":"/tmp/leopold-hookprobe.gjQnXe/codex-project","hook_event_name":"PreToolUse","model":"gpt-5.6-sol","permission_mode":"bypassPermissions","tool_name":"Bash","tool_input":{"command":"true"},"tool_use_id":"exec-c858a358-22fb-46df-92ea-98636929adb5"}

Payload, verbatim (Bash: {"command":"false"}):

{"session_id":"01a064ee-a7c6-7af3-872f-dd265be1acc2","turn_id":"01a064ee-a831-76d0-bfad-138dcf2efcf2","transcript_path":"/private/tmp/leopold-hookprobe.gjQnXe/codex-home/sessions/2026/09/02/rollout-2026-09-02T22-42-37-01a064ee-a7c6-7af3-872f-dd265be1acc2.jsonl","cwd":"/tmp/leopold-hookprobe.gjQnXe/codex-project","hook_event_name":"PreToolUse","model":"gpt-5.6-sol","permission_mode":"bypassPermissions","tool_name":"Bash","tool_input":{"command":"false"},"tool_use_id":"exec-dcaa043b-6d5e-4e48-9fa5-c4aee89bc754"}

Payload, verbatim (Bash: {"command":"grep -c zzz /dev/null"}):

{"session_id":"01a064ee-a7c6-7af3-872f-dd265be1acc2","turn_id":"01a064ee-a831-76d0-bfad-138dcf2efcf2","transcript_path":"/private/tmp/leopold-hookprobe.gjQnXe/codex-home/sessions/2026/09/02/rollout-2026-09-02T22-42-37-01a064ee-a7c6-7af3-872f-dd265be1acc2.jsonl","cwd":"/tmp/leopold-hookprobe.gjQnXe/codex-project","hook_event_name":"PreToolUse","model":"gpt-5.6-sol","permission_mode":"bypassPermissions","tool_name":"Bash","tool_input":{"command":"grep -c zzz /dev/null"},"tool_use_id":"exec-16f28f82-27fc-422d-ac03-8261b9f48361"}

Payload, verbatim (apply_patch: {"command":" Begin Patch\n Update File: /tmp/leopold-hookprobe.gjQnXe/code):

{"session_id":"01a064ee-a7c6-7af3-872f-dd265be1acc2","turn_id":"01a064ee-a831-76d0-bfad-138dcf2efcf2","transcript_path":"/private/tmp/leopold-hookprobe.gjQnXe/codex-home/sessions/2026/09/02/rollout-2026-09-02T22-42-37-01a064ee-a7c6-7af3-872f-dd265be1acc2.jsonl","cwd":"/tmp/leopold-hookprobe.gjQnXe/codex-project","hook_event_name":"PreToolUse","model":"gpt-5.6-sol","permission_mode":"bypassPermissions","tool_name":"apply_patch","tool_input":{"command":"*** Begin Patch\n*** Update File: /tmp/leopold-hookprobe.gjQnXe/codex-project/.leopold/PLAN.md\n@@\n-- [ ] item one\n+- [ ] item one (edited)\n*** End Patch"},"tool_use_id":"exec-c46dd3f8-1a63-4260-94b3-ca57ce5579f6"}

Gatilhos tentados:

• run tools: a headless codex exec turn in a project with an AGENTS.md that runs true, false, grep -c zzz /dev/null, an apply_patch edit of .leopold/PLAN.md, sleep 6 (the driver appends to .leopold/PLAN.md from another process meanwhile) and cat

Respostas que o harness honrou:

• permissionDecision:deny → honrada — PostToolUse fired 0 times after the deny (0 = the tool never ran); reason reached the model: yes (run pretool-deny)
• systemMessage → não observável — marker PROBE_SYSMSG PreToolUse in the --json stream: 0 times; in the rollout transcript: 0 0 times (run system-message)
• additionalContext → não honrada — the model repeated 'PROBE_CTX PreToolUse': no (run additional-context)

PermissionRequest — Codex CLI

Disparou 1 vez(es) na(s) run(s) perm-deny. hook_event_name: present; session_id: present. Chaves do payload: cwd, hook_event_name, model, permission_mode, session_id, tool_input, tool_name, transcript_path, turn_id.

Payload, verbatim:

{"session_id":"01a064f1-6f6f-75e1-bc35-e4fb8d06d50d","turn_id":"01a064f1-6fb2-7720-9805-3bb26d504576","transcript_path":"/private/tmp/leopold-hookprobe.gjQnXe/codex-home/sessions/2026/09/02/rollout-2026-09-02T22-45-39-01a064f1-6f6f-75e1-bc35-e4fb8d06d50d.jsonl","cwd":"/tmp/leopold-hookprobe.gjQnXe/codex-project","hook_event_name":"PermissionRequest","model":"gpt-5.6-sol","permission_mode":"default","tool_name":"Bash","tool_input":{"command":"curl -sI https://example.com | head -1","description":"The sandboxed command returned no output, consistent with curl’s silent mode hiding a network-access failure. Allow this exact command to run once with network access?"}}

Gatilhos tentados:

• run perm-allow: --approve-for-me and a curl that needs network escalation out of the workspace-write sandbox
• run perm-onrequest: -s workspace-write -c approval_policy="on-request" and the same curl (fired: 0 times)

Respostas que o harness honrou:

• decision:allow → não honrada — the escalated curl ran (an HTTP status line came back): no; PermissionRequest fired: 0 (run perm-allow)
• decision:deny → honrada — the deny message reached the model: yes; an HTTP status line came back: no; stderr: Rejected(\"PROBE_DENY\") (run perm-deny)
• decision:allow (approval_policy=on-request) → não honrada — under approval_policy=on-request in headless exec the hook fired 0 times; reply: (run perm-onrequest)

PostToolUse — Codex CLI

Disparou 24 vez(es) na(s) run(s) additional-context, compact, perm-allow, perm-deny, perm-onrequest, posttool-block, subagent, subagent-exit2, subagent-role, system-message, tools. hook_event_name: present; session_id: present. Chaves do payload: cwd, hook_event_name, model, permission_mode, session_id, tool_input, tool_name, tool_response, tool_use_id, transcript_path, turn_id.

Payload, verbatim (Bash: {"command":"true"}):

{"session_id":"01a064ee-a7c6-7af3-872f-dd265be1acc2","turn_id":"01a064ee-a831-76d0-bfad-138dcf2efcf2","transcript_path":"/private/tmp/leopold-hookprobe.gjQnXe/codex-home/sessions/2026/09/02/rollout-2026-09-02T22-42-37-01a064ee-a7c6-7af3-872f-dd265be1acc2.jsonl","cwd":"/tmp/leopold-hookprobe.gjQnXe/codex-project","hook_event_name":"PostToolUse","model":"gpt-5.6-sol","permission_mode":"bypassPermissions","tool_name":"Bash","tool_input":{"command":"true"},"tool_response":"","tool_use_id":"exec-c858a358-22fb-46df-92ea-98636929adb5"}

Payload, verbatim (Bash: {"command":"false"}):

{"session_id":"01a064ee-a7c6-7af3-872f-dd265be1acc2","turn_id":"01a064ee-a831-76d0-bfad-138dcf2efcf2","transcript_path":"/private/tmp/leopold-hookprobe.gjQnXe/codex-home/sessions/2026/09/02/rollout-2026-09-02T22-42-37-01a064ee-a7c6-7af3-872f-dd265be1acc2.jsonl","cwd":"/tmp/leopold-hookprobe.gjQnXe/codex-project","hook_event_name":"PostToolUse","model":"gpt-5.6-sol","permission_mode":"bypassPermissions","tool_name":"Bash","tool_input":{"command":"false"},"tool_response":"","tool_use_id":"exec-dcaa043b-6d5e-4e48-9fa5-c4aee89bc754"}

Payload, verbatim (Bash: {"command":"grep -c zzz /dev/null"}):

{"session_id":"01a064ee-a7c6-7af3-872f-dd265be1acc2","turn_id":"01a064ee-a831-76d0-bfad-138dcf2efcf2","transcript_path":"/private/tmp/leopold-hookprobe.gjQnXe/codex-home/sessions/2026/09/02/rollout-2026-09-02T22-42-37-01a064ee-a7c6-7af3-872f-dd265be1acc2.jsonl","cwd":"/tmp/leopold-hookprobe.gjQnXe/codex-project","hook_event_name":"PostToolUse","model":"gpt-5.6-sol","permission_mode":"bypassPermissions","tool_name":"Bash","tool_input":{"command":"grep -c zzz /dev/null"},"tool_response":"0\n","tool_use_id":"exec-16f28f82-27fc-422d-ac03-8261b9f48361"}

Payload, verbatim (apply_patch: {"command":" Begin Patch\n Update File: /tmp/leopold-hookprobe.gjQnXe/code):

{"session_id":"01a064ee-a7c6-7af3-872f-dd265be1acc2","turn_id":"01a064ee-a831-76d0-bfad-138dcf2efcf2","transcript_path":"/private/tmp/leopold-hookprobe.gjQnXe/codex-home/sessions/2026/09/02/rollout-2026-09-02T22-42-37-01a064ee-a7c6-7af3-872f-dd265be1acc2.jsonl","cwd":"/tmp/leopold-hookprobe.gjQnXe/codex-project","hook_event_name":"PostToolUse","model":"gpt-5.6-sol","permission_mode":"bypassPermissions","tool_name":"apply_patch","tool_input":{"command":"*** Begin Patch\n*** Update File: /tmp/leopold-hookprobe.gjQnXe/codex-project/.leopold/PLAN.md\n@@\n-- [ ] item one\n+- [ ] item one (edited)\n*** End Patch"},"tool_response":"Exit code: 0\nWall time: 0 seconds\nOutput:\nSuccess. Updated the following files:\nM /tmp/leopold-hookprobe.gjQnXe/codex-project/.leopold/PLAN.md\n","tool_use_id":"exec-c46dd3f8-1a63-4260-94b3-ca57ce5579f6"}

Gatilhos tentados:

• run tools: a headless codex exec turn in a project with an AGENTS.md that runs true, false, grep -c zzz /dev/null, an apply_patch edit of .leopold/PLAN.md, sleep 6 (the driver appends to .leopold/PLAN.md from another process meanwhile) and cat

Respostas que o harness honrou:

• observe → n/a — Bash true: tool_response=""; Bash false: PostToolUse fired yes with tool_response="" — no exit code in either; PreToolUse calls without a PostToolUse in the run: 0 (failed apply_patch attempts); edit tool name: apply_patch, (run tools)
• decision:block → honrada — the reason PROBE_BLOCK reached the model/stream: yes; reply: ``` (run posttool-block)
• systemMessage → não observável — marker PROBE_SYSMSG PostToolUse in the --json stream: 0 times; in the rollout transcript: 0 0 times (run system-message)
• additionalContext → honrada — the model repeated 'PROBE_CTX PostToolUse': yes (run additional-context)

PreCompact — Codex CLI

Disparou 1 vez(es) na(s) run(s) compact. hook_event_name: present; session_id: present. Chaves do payload: cwd, hook_event_name, model, session_id, transcript_path, trigger, turn_id.

Payload, verbatim:

{"session_id":"01a064f2-3530-7453-9e93-f1a7093bb39d","turn_id":"01a064f2-3555-7173-9583-c58b54a53f71","transcript_path":"/private/tmp/leopold-hookprobe.gjQnXe/codex-home/sessions/2026/09/02/rollout-2026-09-02T22-46-30-01a064f2-3530-7453-9e93-f1a7093bb39d.jsonl","cwd":"/tmp/leopold-hookprobe.gjQnXe/codex-project","hook_event_name":"PreCompact","model":"gpt-5.6-sol","trigger":"auto"}

Gatilhos tentados:

• run compact: -c model_auto_compact_token_limit=6000 and three full cats of a 120 KB file

Respostas que o harness honrou:

• observe → n/a — auto compaction: PreCompact fired 1 times, PostCompact 1 times (trigger: auto) (run compact)

PostCompact — Codex CLI

Disparou 1 vez(es) na(s) run(s) compact. hook_event_name: present; session_id: present. Chaves do payload: cwd, hook_event_name, model, session_id, transcript_path, trigger, turn_id.

Payload, verbatim:

{"session_id":"01a064f2-3530-7453-9e93-f1a7093bb39d","turn_id":"01a064f2-3555-7173-9583-c58b54a53f71","transcript_path":"/private/tmp/leopold-hookprobe.gjQnXe/codex-home/sessions/2026/09/02/rollout-2026-09-02T22-46-30-01a064f2-3530-7453-9e93-f1a7093bb39d.jsonl","cwd":"/tmp/leopold-hookprobe.gjQnXe/codex-project","hook_event_name":"PostCompact","model":"gpt-5.6-sol","trigger":"auto"}

Gatilhos tentados:

• run compact: -c model_auto_compact_token_limit=6000 and three full cats of a 120 KB file

UserPromptSubmit — Codex CLI

Disparou 24 vez(es) na(s) run(s) additional-context, compact, exec-as-role-4, fail-401, fail-429, fail-500, fail-529, interrupt, perm-allow, perm-deny, perm-onrequest, posttool-block, pretool-deny, prompt-exit2, role-bogus, role-bogus-strict, stop-block, stop-exit2, strict-config, subagent, subagent-exit2, subagent-role, system-message, tools. hook_event_name: present; session_id: present. Chaves do payload: cwd, hook_event_name, model, permission_mode, prompt, session_id, transcript_path, turn_id.

Payload, verbatim:

{"session_id":"01a064ee-a7c6-7af3-872f-dd265be1acc2","turn_id":"01a064ee-a831-76d0-bfad-138dcf2efcf2","transcript_path":"/private/tmp/leopold-hookprobe.gjQnXe/codex-home/sessions/2026/09/02/rollout-2026-09-02T22-42-37-01a064ee-a7c6-7af3-872f-dd265be1acc2.jsonl","cwd":"/tmp/leopold-hookprobe.gjQnXe/codex-project","hook_event_name":"UserPromptSubmit","model":"gpt-5.6-sol","permission_mode":"bypassPermissions","prompt":"Do these steps in order with tools, one tool call per step: 1) run the shell command: true  2) run the shell command: false (it fails; continue)  3) run: grep -c zzz /dev/null (exit 1; continue)  4) edit the file .leopold/PLAN.md with your file-edit tool so 'item one' becomes 'item one (edited)' (keep the leading '- [ ] ')  5) run the shell command: sleep 6  6) run: cat .leopold/PLAN.md. Then reply with the single word DONE."}

Gatilhos tentados:

• run tools: a headless codex exec turn in a project with an AGENTS.md that runs true, false, grep -c zzz /dev/null, an apply_patch edit of .leopold/PLAN.md, sleep 6 (the driver appends to .leopold/PLAN.md from another process meanwhile) and cat

Respostas que o harness honrou:

• exit2 → honrada — the model answered the prompt: no; Stop fired: 0; errors: (run prompt-exit2)
• systemMessage → não observável — marker PROBE_SYSMSG UserPromptSubmit in the --json stream: 0 times; in the rollout transcript: 0 0 times (run system-message)
• additionalContext → não honrada — the model repeated 'PROBE_CTX UserPromptSubmit': no (run additional-context)

Stop — Codex CLI

Disparou 20 vez(es) na(s) run(s) additional-context, compact, exec-as-role-4, perm-allow, perm-deny, perm-onrequest, posttool-block, pretool-deny, role-bogus, role-bogus-strict, stop-block, stop-exit2, strict-config, subagent, subagent-exit2, subagent-role, system-message, tools. hook_event_name: present; session_id: present. Chaves do payload: cwd, hook_event_name, last_assistant_message, model, permission_mode, session_id, stop_hook_active, transcript_path, turn_id.

Payload, verbatim (stop_hook_active: false):

{"session_id":"01a064ee-a7c6-7af3-872f-dd265be1acc2","turn_id":"01a064ee-a831-76d0-bfad-138dcf2efcf2","transcript_path":"/private/tmp/leopold-hookprobe.gjQnXe/codex-home/sessions/2026/09/02/rollout-2026-09-02T22-42-37-01a064ee-a7c6-7af3-872f-dd265be1acc2.jsonl","cwd":"/tmp/leopold-hookprobe.gjQnXe/codex-project","hook_event_name":"Stop","model":"gpt-5.6-sol","permission_mode":"bypassPermissions","stop_hook_active":false,"last_assistant_message":"DONE"}

Payload, verbatim (stop_hook_active: true):

{"session_id":"01a064f0-7715-7743-bfbe-b516cf211385","turn_id":"01a064f0-773e-79d1-b5fc-d8d705d614ca","transcript_path":"/private/tmp/leopold-hookprobe.gjQnXe/codex-home/sessions/2026/09/02/rollout-2026-09-02T22-44-35-01a064f0-7715-7743-bfbe-b516cf211385.jsonl","cwd":"/tmp/leopold-hookprobe.gjQnXe/codex-project","hook_event_name":"Stop","model":"gpt-5.6-sol","permission_mode":"bypassPermissions","stop_hook_active":true,"last_assistant_message":"FIRST_STOP"}

Gatilhos tentados:

• run fail-401: a custom model provider whose base_url is a Python-stdlib stub answering HTTP 401 (the turn fails; which hooks fire?)
• run fail-429: a custom model provider whose base_url is a Python-stdlib stub answering HTTP 429 (the turn fails; which hooks fire?)
• run fail-500: a custom model provider whose base_url is a Python-stdlib stub answering HTTP 500 (the turn fails; which hooks fire?)
• run fail-529: a custom model provider whose base_url is a Python-stdlib stub answering HTTP 529 (the turn fails; which hooks fire?)
• run tools: a headless codex exec turn in a project with an AGENTS.md that runs true, false, grep -c zzz /dev/null, an apply_patch edit of .leopold/PLAN.md, sleep 6 (the driver appends to .leopold/PLAN.md from another process meanwhile) and cat

Respostas que o harness honrou:

• exit2 → honrada — Stop fired 2 times (exit 2 once); stop_hook_active on the second: true; messages: 2 (run stop-exit2)
• decision:block → honrada — Stop fired 2 times ({"decision":"block"} once); stop_hook_active on the second: true (run stop-block)
• systemMessage → não observável — marker PROBE_SYSMSG Stop in the --json stream: 0 times; in the rollout transcript: 0 0 times (run system-message)
• observe (API failure HTTP 401) → n/a — HTTP 401 from the stub (3 request(s): 1 /v1/models; 2 /v1/responses;) → stream: 1 error; 4 item.completed; 1 thread.started; 1 turn.failed; 1 turn.started;; hooks that fired: SessionEnd SessionStart UserPromptSubmit ; error: Your access token could not be refreshed because your refresh token was already used. Please log out and sign in again. Your access token co (run fail-401)
• observe (API failure HTTP 429) → n/a — HTTP 429 from the stub (2 request(s): 1 /v1/models; 1 /v1/responses;) → stream: 1 error; 4 item.completed; 1 thread.started; 1 turn.failed; 1 turn.started;; hooks that fired: SessionEnd SessionStart UserPromptSubmit ; error: exceeded retry limit, last status: 429 Too Many Requests exceeded retry limit, last status: 429 Too Many Requests (run fail-429)
• observe (API failure HTTP 500) → n/a — HTTP 500 from the stub (2 request(s): 1 /v1/models; 1 /v1/responses;) → stream: 1 error; 4 item.completed; 1 thread.started; 1 turn.failed; 1 turn.started;; hooks that fired: SessionEnd SessionStart UserPromptSubmit ; error: We're currently experiencing high demand, which may cause temporary errors. We're currently experiencing high demand, which may cause tempor (run fail-500)
• observe (API failure HTTP 529) → n/a — HTTP 529 from the stub (2 request(s): 1 /v1/models; 1 /v1/responses;) → stream: 1 error; 4 item.completed; 1 thread.started; 1 turn.failed; 1 turn.started;; hooks that fired: SessionEnd SessionStart UserPromptSubmit ; error: unexpected status 529 : PROBE overloaded, url: http://127.0.0.1:19429/v1/responses unexpected status 529 <unknown statu (run fail-529)

Interrupt — Codex CLI

Disparou 1 vez(es) na(s) run(s) interrupt. hook_event_name: present; session_id: present. Chaves do payload: cwd, hook_event_name, model, permission_mode, session_id, transcript_path, turn_id.

Payload, verbatim:

{"session_id":"01a064f2-7877-7f72-8286-b141dc83d8b0","turn_id":"01a064f2-789f-7081-8c7d-352f750ce929","transcript_path":"/private/tmp/leopold-hookprobe.gjQnXe/codex-home/sessions/2026/09/02/rollout-2026-09-02T22-46-47-01a064f2-7877-7f72-8286-b141dc83d8b0.jsonl","cwd":"/tmp/leopold-hookprobe.gjQnXe/codex-project","hook_event_name":"Interrupt","model":"gpt-5.6-sol","permission_mode":"bypassPermissions"}

Gatilhos tentados:

• run interrupt: SIGINT to the codex exec process while its shell command sleeps

Respostas que o harness honrou:

• observe → n/a — after SIGINT: Interrupt fired 1, Stop fired 0, SessionEnd fired 1 (run interrupt)

Conclusões

• Todo evento que uma sessão headless consegue alcançar dispara no Claude Code 2.1.259 (Claude Code), no .claude/settings.json do próprio projeto, com o mesmo prefixo session_id/transcript_path/cwd. Os que não dispararam não têm gatilho headless nesta versão: PermissionDenied (o modo auto nega sem consultá-lo), TeammateIdle (sem ferramenta TeamCreate no -p), DirectoryAdded (/add-dir não existe no -p e o control request register_repo_root responde como registrado), Notification (nada pergunta numa sessão headless) e WorktreeRemove (a saída da sessão --worktree não o chamou) — cada linha abaixo nomeia o gatilho tentado.
• A maioria dos comandos Bash que falham não chega ao PostToolUse no Claude Code — e os que chegam não são sucessos. Exit 0 dispara PostToolUse com tool_response.{stdout,stderr,interrupted} e sem código de saída; exit diferente de zero normalmente dispara PostToolUseFailure, cujo error é a string "Exit code 1" (mais is_interrupt). A exceção está nesta mesma captura: grep -c zzz /dev/null, que sai com 1, chegou ao PostToolUse carregando returnCodeInterpretation: "No matches found" — o harness reinterpreta um conjunto de exits não-zero como "não é erro" e os anota em vez de mandá-los ao evento de falha. Então quem decide é o OBJETO de resposta, não o evento: returnCodeInterpretation (um não-zero reinterpretado), interrupted (cortado no meio) e backgroundTaskId (lançado, não terminado) significam, cada um, que o comando não saiu limpo com 0. O Codex dispara PostToolUse nos dois casos, com tool_response como string e sem código de saída em nenhum deles — um gate de evidência de conclusão precisa ler o resultado do comando do próprio registro, nunca do payload do hook.
• FileChanged não distingue a edição da própria sessão de uma escrita de outro processo, e o matcher faz dois trabalhos. Tanto a ferramenta Edit quanto um append externo produzem o mesmo payload {file_path, event: "change"}; a correlação tem de ser um PostToolUse no mesmo caminho no mesmo instante. O matcher registra um watch como caminho literal relativo à raiz do projeto (PLAN.md, .leopold/PLAN.md) e despacha pelo basename do arquivo alterado (PLAN.md, *) — então um arquivo aninhado precisa de duas entradas, uma em forma de caminho que nunca dispara sozinha e uma em forma de basename que recebe (runs tools, filechanged-nested); PLAN.md|.envrc não registra nada, e um cd na sessão (CwdChanged) desliga todos os watches (run filechanged-cwd). Hooks de FileChanged rodam fora do turno: nunca aparecem como hook_started no stream.
• A ferramenta de edição do Codex é apply_patch (tool_input.command carrega o patch), o shell dele é reportado como Bash, e as ferramentas multi-agente chegam ao PreToolUse como collaborationspawn_agent / collaborationwait_agent. Um arquivo de role em .codex/agents/<role>.toml (name, description, developer_instructions e qualquer chave do config.toml) é honrado por spawn_agent(agent_type=<role>); uma chave desconhecida faz o Codex ignorar o arquivo com um Ignoring malformed agent role definition logado (com e sem --strict-config). codex exec não roda como uma role: agent_role, agent_type e role são rejeitados como campos de config desconhecidos, um arquivo de role não é uma camada --profile válida, e agents.<name>.config_file só declara a role para spawn.
• Exit 2 é honrado onde os docs dizem, nos dois harnesses: TaskCompleted (a task fica pendente), UserPromptSubmit e UserPromptExpansion (o prompt nunca chega ao modelo), Stop / SubagentStop (o agente continua e para de novo com stop_hook_active: true), PreModelSwitch (nenhum PostModelSwitch segue). {"decision":"block"} no Stop faz o mesmo nos dois. O ConfigChange é o único exit 2 que esta página não consegue julgar: um hook de ConfigChange nunca chega ao stream — sem hook_started, sem hook_response, logo sem exit code — e o arquivo manter a edição é verdade tendo o reload sido bloqueado ou adotado, e é por isso que a linha dele diz unobservable. Foi medido à parte, por efeito colateral (uma reescrita dos settings que acrescenta um hook, contra uma run de controle), em ConfigChange exit 2 — Verificação ao vivo: o reload É bloqueado e o arquivo MANTÉM a edição.
• PermissionRequest precisa que exista um caminho de decisão. No Claude Code dispara sob --permission-prompts none (e um CLI -p comum com o host padrão nega sem consultá-lo); decision.behavior allow/deny é honrado (allow roda a ferramenta, deny devolve a mensagem ao modelo). No Codex dispara sob --approve-for-me para uma escalação do sandbox (rede), com o mesmo contrato decision.behavior honrado (Rejected("PROBE_DENY") no stream); approval_policy="on-request" no exec headless nunca pergunta.
• Um erro de API é um StopFailure no Claude Code e um turn.failed simples no Codex. ANTHROPIC_BASE_URL no stub com CLAUDE_CODE_MAX_RETRIES=0 disparou StopFailure para toda classe, com error (não error_type): rate_limit para 429, server_error para 500 e 529, authentication_failed para 401 e para login ausente; Stop não dispara e SessionEnd ainda dispara. O Codex não tem hook de falha: o stub de provider terminou cada turno com turn.failed, e só SessionStart, UserPromptSubmit e SessionEnd dispararam — sem Stop, sem Interrupt.
• Compactação dispara PreCompact/PostCompact nos dois, com trigger (auto / manual); o Claude Code acrescenta custom_instructions antes e o compact_summary completo depois, e então dispara SessionStart com source: "compact"; o stream carrega um compact_boundary com pre_tokens/post_tokens. Os payloads do Codex carregam turn_id e model e nada sobre o resumo.
• Interrupt é real no Codex: SIGINT no codex exec no meio de um comando o dispara (sem Stop), depois SessionEnd. O Codex limita os timeouts dos hooks de SessionEnd e Interrupt a 3 s (clamping … hook timeout to 3s no stream) — um hook nesses eventos precisa ser rápido.
• O processo do hook herda o ambiente do pai nos dois harnesses (CODEX_HOME chegou ao hook; o Claude Code acrescenta CLAUDE_CODE_SESSION_ID, CLAUDE_PROJECT_DIR, CLAUDE_ENV_FILE, CLAUDE_CODE_ENTRYPOINT; o Codex não acrescenta nada seu), e hook_event_name e session_id estavam presentes em todo payload capturado dos dois harnesses — um hook pode confiar neles, e o fallback nome-por-argumento do dump hook nunca precisou agir.

O que isto decide para o Leopold

• Toda capacidade da missão desta run é plugada só em eventos capturados acima, com os campos de payload que as capturas mostram; um campo que o probe não capturou manda o item de volta ao probe, nunca a um palpite.
• Onde o Codex não tem o evento (StopFailure, PostToolUseFailure, TaskCompleted, FileChanged, ConfigChange), a capacidade sai no Claude Code com o substituto do Codex que as capturas permitem — PostToolUse mais a edição do PLAN.md para evidência de conclusão, turn.failed no stream --json para erros de API — e o leopold doctor nomeia o que o Codex não consegue, citando a versão que esta página registra.
• Os hooks de Interrupt e SessionEnd no Codex ficam abaixo de 3 s; um hook que precise fazer mais registra a intenção e deixa o turno seguinte terminar.

Runs

Harness	Run	rc	s	Comando
claude	tools	0	38	claude -p --setting-sources project --strict-mcp-config --model haiku --output-format stream-json --include-hook-events --verbose --permission-mode bypassPermis
claude	filechanged-cwd	0	30	claude -p --setting-sources project --strict-mcp-config --model haiku --output-format stream-json --include-hook-events --verbose --permission-mode bypassPermis
claude	filechanged-nested	0	16	claude -p --setting-sources project --strict-mcp-config --model haiku --output-format stream-json --include-hook-events --verbose --permission-mode bypassPermis
claude	tasks	0	11	claude -p --setting-sources project --strict-mcp-config --model haiku --output-format stream-json --include-hook-events --verbose --permission-mode bypassPermis
claude	tasks-exit2	0	11	claude -p --setting-sources project --strict-mcp-config --model haiku --output-format stream-json --include-hook-events --verbose --permission-mode bypassPermis
claude	subagent	0	11	claude -p --setting-sources project --strict-mcp-config --model haiku --output-format stream-json --include-hook-events --verbose --permission-mode bypassPermis
claude	subagent-exit2	0	14	claude -p --setting-sources project --strict-mcp-config --model haiku --output-format stream-json --include-hook-events --verbose --permission-mode bypassPermis
claude	config-change	0	9	claude -p --setting-sources project --strict-mcp-config --model haiku --output-format stream-json --include-hook-events --verbose --permission-mode bypassPermis
claude	config-change-exit2	0	8	claude -p --setting-sources project --strict-mcp-config --model haiku --output-format stream-json --include-hook-events --verbose --permission-mode bypassPermis
claude	command	0	2	claude -p --setting-sources project --strict-mcp-config --model haiku --output-format stream-json --include-hook-events --verbose --permission-mode bypassPermis
claude	command-exit2	0	1	claude -p --setting-sources project --strict-mcp-config --model haiku --output-format stream-json --include-hook-events --verbose --permission-mode bypassPermis
claude	prompt-exit2	0	1	claude -p --setting-sources project --strict-mcp-config --model haiku --output-format stream-json --include-hook-events --verbose --permission-mode bypassPermis
claude	stop-exit2	0	7	claude -p --setting-sources project --strict-mcp-config --model haiku --output-format stream-json --include-hook-events --verbose --permission-mode bypassPermis
claude	stop-block	0	5	claude -p --setting-sources project --strict-mcp-config --model haiku --output-format stream-json --include-hook-events --verbose --permission-mode bypassPermis
claude	posttool-block	0	6	claude -p --setting-sources project --strict-mcp-config --model haiku --output-format stream-json --include-hook-events --verbose --permission-mode bypassPermis
claude	pretool-deny	0	5	claude -p --setting-sources project --strict-mcp-config --model haiku --output-format stream-json --include-hook-events --verbose --permission-mode bypassPermis
claude	perm-allow	0	6	claude -p --setting-sources project --strict-mcp-config --model haiku --output-format stream-json --include-hook-events --verbose --permission-prompts none
claude	perm-deny	0	6	claude -p --setting-sources project --strict-mcp-config --model haiku --output-format stream-json --include-hook-events --verbose --permission-prompts none
claude	perm-host	0	7	claude -p --setting-sources project --strict-mcp-config --model haiku --output-format stream-json --include-hook-events --verbose
claude	perm-auto	0	8	claude -p --setting-sources project --strict-mcp-config --model haiku --output-format stream-json --include-hook-events --verbose --permission-mode auto
claude	system-message	0	4	claude -p --setting-sources project --strict-mcp-config --model haiku --output-format stream-json --include-hook-events --verbose --permission-mode bypassPermis
claude	additional-context	0	8	claude -p --setting-sources project --strict-mcp-config --model haiku --output-format stream-json --include-hook-events --verbose --permission-mode bypassPermis
claude	compact-auto	0	37	claude -p --setting-sources project --strict-mcp-config --model haiku --output-format stream-json --include-hook-events --verbose --permission-mode bypassPermis
claude	compact-manual-seed	0	6	claude -p --setting-sources project --strict-mcp-config --model haiku --output-format stream-json --include-hook-events --verbose --permission-mode bypassPermis
claude	compact-manual	0	19	claude -p --setting-sources project --strict-mcp-config --model haiku --output-format stream-json --include-hook-events --verbose --permission-mode bypassPermis
claude	fail-401	1	0	claude -p --setting-sources project --strict-mcp-config --model haiku --output-format stream-json --include-hook-events --verbose --permission-mode bypassPermis
claude	fail-429	1	1	claude -p --setting-sources project --strict-mcp-config --model haiku --output-format stream-json --include-hook-events --verbose --permission-mode bypassPermis
claude	fail-500	1	1	claude -p --setting-sources project --strict-mcp-config --model haiku --output-format stream-json --include-hook-events --verbose --permission-mode bypassPermis
claude	fail-529	1	0	claude -p --setting-sources project --strict-mcp-config --model haiku --output-format stream-json --include-hook-events --verbose --permission-mode bypassPermis
claude	elicit-observe	1	0	claude -p --setting-sources project --strict-mcp-config --model haiku --output-format stream-json --include-hook-events --verbose --permission-mode bypassPermis
claude	elicit-accept	1	0	claude -p --setting-sources project --strict-mcp-config --model haiku --output-format stream-json --include-hook-events --verbose --permission-mode bypassPermis
claude	worktree	0	4	claude -p --setting-sources project --strict-mcp-config --model haiku --output-format stream-json --include-hook-events --verbose --permission-mode bypassPermis
claude	worktree-agent	0	10	claude -p --setting-sources project --strict-mcp-config --model haiku --output-format stream-json --include-hook-events --verbose --permission-mode bypassPermis
claude	setup-init	0	3	claude -p --setting-sources project --strict-mcp-config --model haiku --output-format stream-json --include-hook-events --verbose --permission-mode bypassPermis
claude	setup-maintenance	0	3	claude -p --setting-sources project --strict-mcp-config --model haiku --output-format stream-json --include-hook-events --verbose --permission-mode bypassPermis
claude	model-switch	0	1	claude -p --setting-sources project --strict-mcp-config --model haiku --output-format stream-json --include-hook-events --verbose --permission-mode bypassPermis
claude	model-switch-exit2	0	1	claude -p --setting-sources project --strict-mcp-config --model haiku --output-format stream-json --include-hook-events --verbose --permission-mode bypassPermis
claude	add-dir	0	1	claude -p --setting-sources project --strict-mcp-config --model haiku --output-format stream-json --include-hook-events --verbose --permission-mode bypassPermis
claude	add-dir-control	0	0	claude -p --input-format stream-json (control_request register_repo_root directory=/tmp/leopold-hookprobe.gjQnXe/extra-dir)
claude	teams	0	5	claude -p --setting-sources project --strict-mcp-config --model haiku --output-format stream-json --include-hook-events --verbose --permission-mode bypassPermis
claude	unauth	1	0	CLAUDE_CONFIG_DIR=<temp> claude -p --setting-sources project --model haiku --permission-mode bypassPermissions
codex	tools	0	40	codex exec --json --skip-git-repo-check --dangerously-bypass-hook-trust -C <project> --dangerously-bypass-approvals-and-sandbox
codex	subagent	0	19	codex exec --json --skip-git-repo-check --dangerously-bypass-hook-trust -C <project> --dangerously-bypass-approvals-and-sandbox
codex	subagent-role	0	16	codex exec --json --skip-git-repo-check --dangerously-bypass-hook-trust -C <project> --dangerously-bypass-approvals-and-sandbox
codex	role-bogus	0	10	codex exec --json --skip-git-repo-check --dangerously-bypass-hook-trust -C <project> --dangerously-bypass-approvals-and-sandbox
codex	role-bogus-strict	0	11	codex exec --json --skip-git-repo-check --dangerously-bypass-hook-trust -C <project> --dangerously-bypass-approvals-and-sandbox --strict-config
codex	subagent-exit2	0	21	codex exec --json --skip-git-repo-check --dangerously-bypass-hook-trust -C <project> --dangerously-bypass-approvals-and-sandbox
codex	stop-exit2	0	12	codex exec --json --skip-git-repo-check --dangerously-bypass-hook-trust -C <project> --dangerously-bypass-approvals-and-sandbox
codex	stop-block	0	12	codex exec --json --skip-git-repo-check --dangerously-bypass-hook-trust -C <project> --dangerously-bypass-approvals-and-sandbox
codex	prompt-exit2	0	4	codex exec --json --skip-git-repo-check --dangerously-bypass-hook-trust -C <project> --dangerously-bypass-approvals-and-sandbox
codex	posttool-block	0	12	codex exec --json --skip-git-repo-check --dangerously-bypass-hook-trust -C <project> --dangerously-bypass-approvals-and-sandbox
codex	pretool-deny	0	11	codex exec --json --skip-git-repo-check --dangerously-bypass-hook-trust -C <project> --dangerously-bypass-approvals-and-sandbox
codex	perm-allow	0	12	codex exec --json --skip-git-repo-check --dangerously-bypass-hook-trust -C <project> --approve-for-me
codex	perm-deny	0	18	codex exec --json --skip-git-repo-check --dangerously-bypass-hook-trust -C <project> --approve-for-me
codex	perm-onrequest	0	12	codex exec --json --skip-git-repo-check --dangerously-bypass-hook-trust -C <project> -s workspace-write -c approval_policy="on-request"
codex	system-message	0	8	codex exec --json --skip-git-repo-check --dangerously-bypass-hook-trust -C <project> --dangerously-bypass-approvals-and-sandbox
codex	additional-context	0	12	codex exec --json --skip-git-repo-check --dangerously-bypass-hook-trust -C <project> --dangerously-bypass-approvals-and-sandbox
codex	compact	0	16	codex exec --json --skip-git-repo-check --dangerously-bypass-hook-trust -C <project> --dangerously-bypass-approvals-and-sandbox -c model_auto_compact_token_limi
codex	interrupt	1	8	codex exec … (SIGINT sent to the process after PreToolUse of sleep 20)
codex	fail-401	1	2	codex exec --json --skip-git-repo-check --dangerously-bypass-hook-trust -C <project> --dangerously-bypass-approvals-and-sandbox -c model_provider="probe" -c mod
codex	fail-429	1	1	codex exec --json --skip-git-repo-check --dangerously-bypass-hook-trust -C <project> --dangerously-bypass-approvals-and-sandbox -c model_provider="probe" -c mod
codex	fail-500	1	2	codex exec --json --skip-git-repo-check --dangerously-bypass-hook-trust -C <project> --dangerously-bypass-approvals-and-sandbox -c model_provider="probe" -c mod
codex	fail-529	1	1	codex exec --json --skip-git-repo-check --dangerously-bypass-hook-trust -C <project> --dangerously-bypass-approvals-and-sandbox -c model_provider="probe" -c mod
codex	exec-as-role-1	1	0	codex exec --strict-config -c agent_role="probe-reviewer"
codex	exec-as-role-2	1	0	codex exec --strict-config -c agent_type="probe-reviewer"
codex	exec-as-role-3	1	0	codex exec --strict-config -c role="probe-reviewer"
codex	exec-as-role-4	0	0	codex exec --strict-config -c agents.probe-reviewer.config_file=".codex/agents/probe-reviewer.toml"
codex	exec-as-role-profile	1	0	codex exec --strict-config -p probe-reviewer (the role file copied as CODEX_HOME/probe-reviewer.config.toml)
codex	strict-config	0	6	codex exec --json --skip-git-repo-check --dangerously-bypass-hook-trust -C <project> --dangerously-bypass-approvals-and-sandbox --strict-config

Hermeticidade

Toda escrita caiu sob a raiz mktemp -d (/tmp/leopold-hookprobe.gjQnXe), sob <out>, ou nos diretórios de transcript dos próprios harnesses para aquele projeto temporário. Os fingerprints de nomes de entrada de ~/.claude e ~/.codex reais foram idênticos antes e depois (inalterado / MUDOU); o auth.json copiado para o CODEX_HOME temporário foi apagado ao fim do probe. Para repetir: bash scripts/probe-hook-events.sh --out <dir> (nunca faz parte do make test; make hooks-check faz o lint).