Use to toggle the safe-mode flag at .claude/safe-mode. The flag is written automatically by consecutive-failure-guard once tool failures hit FORGE_SAFE_MODE_THRESHOLD (default 5), and block-destructive.sh denies every Bash/Write/Edit while it exists. /safe-mode off clears the flag, logs the exit, and prompts /postmortem; /safe-mode on enters manually; /safe-mode status reports current state.
설치
Codex 또는 Claude로 설치 이 Prompt를 복사해 Codex, Claude 또는 다른 어시스턴트에 붙여 넣으면 Skill 페이지를 검토하고 설치를 진행할 수 있습니다.
Use to toggle the safe-mode flag at .claude/safe-mode. The flag is written automatically by consecutive-failure-guard once tool failures hit FORGE_SAFE_MODE_THRESHOLD (default 5), and block-destructive.sh denies every Bash/Write/Edit while it exists. /safe-mode off clears the flag, logs the exit, and prompts /postmortem; /safe-mode on enters manually; /safe-mode status reports current state.
when_to_use
Reach for "/safe-mode off" after diagnosing the failure chain that auto-triggered the lock, or "/safe-mode on" when about to perform a risky operation that you want the harness to block until you explicitly clear it. Do NOT use to skip a postmortem — clearing the flag without root-cause analysis defeats the graceful-degradation contract. Do NOT use for routine harness health checks — use `/healthcheck` instead; safe-mode only toggles the destructive-edit lockdown.
disable-model-invocation
true
argument-hint
<on|off|status>
allowed-tools
["Read","Write","Bash"]
logical
.claude/safe-mode flag toggled to the requested state; ledger entry appended
/safe-mode — Graceful-Degradation Toggle
Controls .claude/safe-mode. When present, block-destructive.sh denies Bash/Write/Edit. Exists to force a human checkpoint after repeated failures, per TRAE §5.2.4 "Graceful Degradation".
Subcommands
/safe-mode status
Read the flag. Print the four base fields, then — when .brief_template is a non-null object — emit a populated 4-field brief using its values, per ## Escalation Brief below. jq -r '.brief_template.trigger // empty' returns the pre-filled trigger string for auto-written flags; manual writes seed .context / .recommendation placeholders that the agent then fills before printing.
Manual entry. Write the flag with the same shape consecutive-failure-guard.sh uses, so /safe-mode status and any future reader can treat auto and manual entries uniformly:
{"entered_at":"<UTC>","reason":"<reason or 'manual'>","counter":0,"last_tool":"manual","brief_template":{"context":"<one line — what the agent is about to do>","trigger":"Manual entry — <one line reason>","options":["<option, agent can do this without help>","<option, requires your action>","<option, requires your action>"],"recommendation":"<option #N>. <one-line reason>"}}
Suggest /postmortem on the failure chain — the whole point of the forced checkpoint is to learn, not to skip past it:
Safe-mode cleared. The failure chain that triggered it is still worth understanding.
Recommend: run /postmortem on the root cause before continuing.
Execution Checklist
/safe-mode status — read .claude/safe-mode; print the four base fields (state, Entered, Reason, Failure counter) and, when .brief_template is non-null, render its populated 4-field Escalation Brief
/safe-mode on [reason] — write .claude/safe-mode with the consecutive-failure-guard.sh JSON shape (entered_at, reason, counter, last_tool, brief_template); append safe-mode-enter ledger entry; report Safe-mode active. Mutations blocked until /safe-mode off.
/safe-mode off — refuse if no flag; delete .claude/safe-mode; reset /tmp/claude-failure-guard/<session>/consecutive-failures; append safe-mode-exit ledger entry; suggest /postmortem on the failure chain
When activating (auto or manual), emit the Escalation Brief in the documented CONTEXT / TRIGGER / OPTIONS / RECOMMENDATION shape so a human can decide fast
Integration
Writer:plugins/context-engine/hooks/consecutive-failure-guard.sh writes the flag at FORGE_SAFE_MODE_THRESHOLD (default 5).
Ledger: All entries use the SEPL schema so /rest-audit Reliability axis and /lineage-audit can inspect them.
Partner skill:/postmortem (evaluator plugin) — the intended follow-up after every exit.
Known Failure Modes
.claude/safe-mode exists but malformed JSON → treat as active (safer default); /safe-mode off still clears it.
Ledger write fails → flag state still changes (the flag is the source of truth; ledger is audit metadata).
Do NOT
Do not clear the flag silently or auto-reset it. The human checkpoint is the value.
Do not extend safe-mode with new blocking scopes in this skill — that belongs in block-destructive.sh, which is the enforcement layer.
Escalation Brief
When this skill activates — automatically via consecutive-failure-guard, or manually via /safe-mode on — emit a structured brief in this exact shape so a human can decide fast. Reproduce the labels verbatim; a downstream parser detects a brief by them.
CONTEXT: <one line — what was the agent trying to do>
TRIGGER: <one line — what stopped it: failure-count threshold, low-confidence signals, denied permission, conflicting evidence>
OPTIONS:
1. <option, agent can do this without help>
2. <option, requires your action>
3. <option, requires your action>
RECOMMENDATION: <option #N>. <one-line reason>
Waiting for instruction. <thing the agent is paused on> is not blocked yet.
The brief is not a substitute for /postmortem after /safe-mode off — it precedes the human decision; postmortem follows the resolution.
Input: consecutive-failure-guard wrote .claude/safe-mode after 5 failed pytest -k user_signup runs; the agent invokes /safe-mode status.
Output:
SAFE MODE: active
Entered: 2026-05-16T14:22:01Z
Reason: consecutive-failures
Failure counter: 5
CONTEXT: Reproduce the user_signup test failure flagged by /verify.
TRIGGER: Failure-count threshold (5 consecutive pytest exits, last was assertion error in fixture teardown).
OPTIONS:
1. Roll back the last commit and re-run from a known-green tree.
2. Show you the failing fixture and the last three diff hunks for a manual decision.
3. Skip user_signup and gate only on the other 41 tests for now.
RECOMMENDATION: Option 2. The failure is in fixture teardown, not the assertion under test — a human glance at the diff is cheaper than another speculative rewrite.
Waiting for instruction. The working tree is not blocked yet — no further edits will be attempted until you reply.
Input: the agent invokes /safe-mode on risky-migration before running a one-shot database schema change it cannot easily undo.
Output:
Safe-mode active. Mutations blocked until /safe-mode off.
CONTEXT: Apply the `add_audit_columns` migration to the production-shape staging DB before the release window.
TRIGGER: Manual entry — operation is one-shot and the rollback path requires a snapshot restore.
OPTIONS:
1. Run the migration with `--dry-run` first and report the planned DDL.
2. Pause for you to confirm the migration plan and the rollback steps you want recorded.
3. Hand off the migration entirely — I draft the SQL, you run it.
RECOMMENDATION: Option 1. A dry-run is reversible and surfaces any unexpected DDL before the destructive write.
Waiting for instruction. The migration is not blocked yet — the connection is open and the dry-run command is queued.
