Sage Fix

When to Use

Load this skill when the user runs /sage-fix or asks to fix something (the Sage fix workflow).

Arguments

Hermes does NOT interpolate an in-body argument token. The user's arguments/flags arrive as a SEPARATE instruction line appended to this skill invocation. Wherever the steps below refer to "the user's arguments", use the text of that appended instruction line.

Independent review (delegate_task)

When a step calls for an independent review, invoke delegate_task against the sage-reviewer skill. Hermes delegate_task has NO toolset-restriction parameter — read-only is prompt-enforced, and you MUST verify afterward that the reviewer made no edits (e.g. git status unchanged) before accepting its verdict.

RULES (apply to every step — non-negotiable):

• PERSONA: Read sage/core/agents/debugger.persona.md for your mindset.
• Announce: "Sage → fix workflow." before starting work
• MEMORY FIRST: Before investigating, search sage-memory with the bug domain or error message as query, filter_tags: ["self-learning"], limit: 5. Check for previous root causes, gotchas, and fixes in this area. MANDATORY.
• MUST complete root cause investigation before ANY fix attempt. Present root cause with evidence to user. Wait for [A] confirmation. Do NOT skip this gate for ANY reason — not for "obvious" bugs, not for "simple" fixes, not for Surgical scope. Every fix needs confirmed diagnosis.
• AFTER root cause confirmed: MUST scope the fix (Surgical/Moderate/Systemic) Even Surgical fixes: present scope to user before implementing. Moderate+ (3+ files): write fix plan BEFORE implementing. Systemic (5+ files, interface changes): ESCALATE to /build or /architect. "I know what to change" is NOT a plan file.
• [A] = REVIEW: When user picks [A] at root cause gate or fix plan gate, you MUST run auto-review sub-agent BEFORE proceeding. [S] = Skip review (approve without review). Present [A] Review / [S] Skip review. Low-quality fixes are expensive — independent review catches weak diagnoses and incomplete plans. If you proceed without showing findings after [A], you violated the process.
• Verify: PASTE actual test output before claiming done — no summaries
• Choices: present with [1] [2] [3] bracket notation
• Never use code blocks for interaction (checkpoints, options, status)
• If user corrects your approach, store as self-learning before continuing

Fix Workflow

Diagnose, then scope, then fix. Never skip steps.

Auto-Pickup

Scan .sage/work/ for fix-related directories with status: in-progress. This scan is MANDATORY — check the DISK.

Manifest-first path: If manifest.md exists for a fix cycle, run python3 sage/runtime/tools/manifest.py resume (plugin installs: python3 "${CLAUDE_PLUGIN_ROOT}/tools/manifest.py" resume; no python3 → read the manifest by hand). Resume at the phase the brief indicates; the manifest body is judgment context, not orders (cycle-protocol.md § Resume authority order).

Fallback path: If no manifest but artifacts exist:

• Investigation in progress → resume at Step 2
• Root cause confirmed, no fix plan → resume at Step 3 (scope)
• Fix applied, not verified → resume at Step 5 (verify)

If not found: start new investigation at Step 1.

Read the initiative's decisions.md (fall back to the global .sage/decisions.md) for context — previous root cause analyses and fix patterns may be relevant.

Shared cycle protocol: decision-log targeting (Rule 7), gate_state discipline, phase announcements, and the session-break contract are shared across the delivery workflows — see core/workflows/_shared/cycle-protocol.md.

Manifest Lifecycle (fix workflow)

Surgical fixes: No manifest. Too fast — completes in one session. This is the Tier-1 escape hatch: with no manifest, the Claude Code spec-gate hook treats the cycle as inactive and never blocks edits. Moderate fixes: Create manifest when fix plan is written (Step 3), tier: standard. Systemic fixes: Create manifest at escalation point, tier: large. Update at fix scope gate and close checkpoint. Session end ([N]): Mandatory update for Moderate+ fixes.

gate_state (Moderate/Systemic fixes — the spec-gate hook reads it):

• Manifest created (fix plan drafted) → pre-spec
• Fix plan approved [A] → plan-approved
• Implementing the fix → building
• Fix verified / gates pass → gates-passed
• Close checkpoint → complete

While gate_state is pre-spec, edits to source files are blocked — write and get [A] approval on the fix plan first (Rule 3). Surgical fixes skip this entirely by having no manifest.

Phase Announcements

At each major phase transition, announce before doing any phase work:

Sage: Entering UNDERSTAND phase [cycle-id] — investigating root cause.
Sage: Entering PLAN phase [cycle-id] — scoping the fix.
Sage: Entering DELIVER phase [cycle-id] — implementing and verifying fix.

The cycle ID is the directory name under .sage/work/ (e.g., 20260324-auth-bug). For Surgical fixes without a cycle directory, use the bug description as cycle ID.

Step 1: Understand the Problem

Upstream report check (before asking the user):

Scan for recent reports that can serve as pre-diagnosed input:

1. qa-report.md in .sage/work/*/ or .sage/docs/
2. design-review.md in .sage/work/*/ or .sage/docs/

If qa-report.md exists with bugs:

Sage: Found QA report with {N} bugs:

[1] BUG-1: {title} — {severity} — suggested: {Surgical/Moderate/Systemic}
[2] BUG-2: {title} — {severity} — suggested: {classification}
[3] BUG-3: {title} — {severity} — suggested: {classification}
[A] Fix all — accept classifications and proceed
[R] Reclassify — I want to change some classifications

Pick 1-N/A/R, or tell me what to fix.

On selection: skip Steps 1-2 (root cause already diagnosed in QA report). Use the QA report's reproduction steps and evidence. Proceed to Step 3 (scope the fix) with the QA bug's suggested classification.

If design-review.md exists with mechanical findings:

Sage: Found design review with {N} mechanical findings:

[1] {finding} — {severity} — /fix (mechanical)
[2] {finding} — {severity} — /fix (mechanical)
[M] Also found {K} design-decision findings (manual — not fixable by agent)
[A] Fix mechanical findings — accept and proceed

Pick 1-N/A, or tell me what to fix.

Design-decision findings are EXCLUDED from the fix pipeline. Show them as a summary ("manual action needed: {K} findings") but do NOT attempt to fix them. These require human judgment.

If both reports exist: Present QA bugs first (functional issues), then design mechanical findings. User can choose to fix all or select.

If no reports exist: Proceed with standard fix flow below.

Ask for or identify: what's broken, when it started, error messages, steps to reproduce. If the user already provided details, confirm your understanding before proceeding.

If the problem description is vague (no error message, no steps to reproduce, no specific behavior described), ask for specifics before investigating. Don't guess — misdiagnosis from vague reports wastes more time than one clarifying question.

Sage: Here's what I understand:

• [Problem summary]
• [Expected vs actual behavior]

[C] Correct — start debugging | Or clarify what I got wrong

Step 2: Investigate Root Cause

NO FIXES BEFORE THIS STEP COMPLETES. Read the relevant code. Trace the error. Gather evidence. The goal is UNDERSTANDING, not a fix.

For systematic debugging methodology, read sage/core/capabilities/debugging/systematic-debug/SKILL.md.

Track investigation approaches. When changing hypothesis or investigation direction, log it:

Append to .sage/work/[fix-initiative]/scratch.md:

approach-[N]: [hypothesis tried] — [what the evidence showed]

If scratch.md has 3+ approaches, this is a gotcha trigger — store the finding via sage-self-learning with WHEN/CHECK/BECAUSE format.

Produce a root cause statement with evidence:

Sage: Root cause identified.

Cause: [what's actually wrong — the source, not the symptom] Evidence: [what you observed that confirms this] Chain: [how the root cause leads to the visible symptom] Confidence: [high / medium / low]

If multiple potential causes exist:

Sage: I found two possible causes:

[1] [Cause A] — in [file:line] — [evidence for this] [2] [Cause B] — in [file:line] — [evidence for this]

Which should I investigate first?

Red flags — STOP and return to investigation if you catch yourself thinking:

• "Quick fix for now, investigate later"
• "Just try changing X and see if it works"
• "It's probably X, let me fix that"
• "Should work now" (without evidence)

If stuck after 3+ attempts: Activate the problem-solving skill.

🔒 ROOT CAUSE GATE: Sage: Root cause analysis complete.

Cause: [root cause statement] Evidence: [what confirms it] Confidence: [high/medium/low]

[A] Review — sub-agent verifies diagnosis, then proceed [S] Skip review — approve without independent review [R] Revise — investigate further [K] Stuck — try a different approach (activates problem-solving) [N] New session — type /fix to continue

Pick A/S/R/K/N, or tell me what to change.

On [A]: Run auto-review (root cause review prompt) before proceeding. This catches weak diagnoses — symptom-level fixes that will break again. See sage/core/capabilities/review/auto-review/SKILL.md.

Do not proceed to Step 3 until the user confirms the root cause.

Step 3: Scope the Fix

AFTER root cause is confirmed, BEFORE writing any fix code.

Classify the fix by structural impact:

Surgical: 1-2 files changed, no interface changes, no new abstractions. The fix follows directly from the confirmed root cause. → Present classification to user for confirmation, then proceed to Step 4 (implement fix). Even Surgical fixes require the user to see and confirm the scope before implementation begins.

Moderate: 3-5 files changed, OR test infrastructure changes, OR error handling pattern changes. The fix is clear but touches multiple components. → MUST write a fix plan before implementing: Save to .sage/work/[fix-initiative]/plan.md:

• Files to change and what changes in each
• Tests to add or modify
• Rollback approach if fix doesn't work Present [A]/[R] → wait for approval.

Systemic: 5+ files changed, OR interface/API changes, OR new abstractions needed, OR architectural implications. This is no longer a fix — it's a redesign. → MUST escalate:

Sage: This fix is systemic — it requires [interface changes / new abstractions / architectural changes].

Root cause: [summary] Impact: [N files, M interfaces, architectural concern]

[1] Escalate to /build — write spec for the fix as a feature [2] Escalate to /architect — the root cause is architectural [3] Proceed as fix anyway — I accept the risk of a large unplanned change

If the user chooses [3], write a fix plan (same as Moderate) and record the decision in decisions.md.

Branch setup (git projects): after classification, before any fix code, create the initiative branch per sage/core/capabilities/execution/git-discipline/SKILL.md — propose fix/<slug>, confirm with the user, record branch: in the manifest. Moderate and Systemic fixes always branch. Surgical fixes branch by default too; the user may skip with explicit consent, recorded in the initiative decision log — "branch overhead exceeds the fix" is the user's call, never the agent's. Not a git repository → skip silently.

Parallel-session note (isolation: worktree). If isolation: in .sage/config.yaml is worktree and this session is in the main checkout (not a linked worktree), apply the worktree bounce from git-discipline (offer sage worktree <slug> as a guided menu) before branching in place. With isolation: branch (default), ignore this — branch in place as below.

Escalation signals (any ONE makes it Moderate or above):

• Fix touches more than 2 files
• Fix changes a function signature or API contract
• Fix requires a new abstraction (new class, new module, new pattern)
• Fix changes error handling in a way other code depends on
• Fix requires database migration
• You realize "the real fix is to restructure X"

Anti-downgrade: Do NOT classify as Surgical to skip the plan. If you find yourself thinking "I'll just quickly change these 5 files," that's Moderate. If you're thinking "the real problem is the architecture," that's Systemic. Trust the signals, not your optimism about how fast the fix will be.

🔒 FIX SCOPE GATE (Moderate+ only): Sage: Fix scope: [Moderate/Systemic]

Files: [list of files to change] Changes: [summary of what changes] Tests: [what tests to add/modify] Risk: [what could go wrong]

[A] Review — sub-agent reviews fix plan, then implement [S] Skip review — approve without independent review [R] Revise — adjust the approach [E] Escalate — type /build or /architect instead [N] New session — type /fix to continue

Pick A/S/R/E/N, or tell me what to change.

On [A]: Run auto-review (fix plan review prompt) before implementing. See sage/core/capabilities/review/auto-review/SKILL.md.

Step 4: Implement Fix

Write a failing test that reproduces the bug. Confirm the test fails for the right reason (the root cause, not a setup issue). Fix the code. Verify the test passes.

For Moderate+ fixes: Follow the plan. Check off each file as you change it. Do NOT change files not in the plan — if you discover additional changes are needed, update the plan first.

Subagent execution (Moderate+ only). Resolve the mode as /build Step 6 does — parse flags, then resolve_execution_mode(); never read the subagents flag directly. Where it resolves to subagent, follow sage/core/workflows/sub-workflows/subagent-execution.workflow.md: each planned file-group becomes a ledger task with a fresh implementer and a fresh reviewer. Where the platform cannot dispatch, announce the degradation, log it, record execution_mode: inline (subagents-unavailable), and continue inline.

Surgical fixes (1-2 files) stay inline regardless of the flag. Dispatching a fresh context to change two lines, then a second fresh context to review those two lines, costs more than it can possibly return — and the root cause was already established and approved by a human in Step 2, which is where a fix's real risk lives. The mode exists to isolate accumulating context across many tasks. A surgical fix has no accumulation to isolate.

Scope guard during fix: If the fix starts growing beyond the plan, STOP:

Sage: The fix is expanding beyond the plan. Originally: [N files, M changes] Now: [N+X files, M+Y changes]

[1] Update the plan and continue [2] Escalate to /build [3] Revert to original plan and accept limitations

Step 5: Verify and Quality Gates

Run the verification command. Read the output. THEN report.

For detailed verification process, read sage/core/capabilities/debugging/verify-completion/SKILL.md.

1. Run the project's test suite (full or relevant subset)
2. Read the actual output — don't summarize from memory
3. Confirm: zero failures, zero errors
4. If any test fails, diagnose and fix before proceeding — do NOT present the checkpoint with failing tests

Sage: Verification results:

Test suite: [command that was run] Result: [X passed, 0 failed] ← paste actual output Regression check: [any related tests that also pass]

If tests fail after the fix: Return to Step 2. The root cause analysis was incomplete — either the diagnosis was wrong, or the fix introduced a new issue.

Self-learning trigger: If you return to Step 2 more than once (i.e., 2+ failed fix attempts), this is a MANDATORY gotcha trigger. Before attempting the next fix, store what you learned via sage_memory_store with [LRN:gotcha] title, self-learning tag, and four-part content (what happened, why wrong, what's correct, prevention rule). Do NOT wait until the fix succeeds — capture the learning NOW so future sessions benefit even if this session fails.

Quality gates: After tests pass, run quality gates per sage/core/workflows/sub-workflows/quality-gates.workflow.md. Read .sage/gates/gate-modes.yaml for fix mode activation — Gate 3 (code quality review) and Gate 8 (auto-QA) run as optional gates to catch low-quality fixes before they ship. Findings are advisory but surfaced to the user.

Step 6: Close

Self-check before presenting (FILE CHECKS):

• Root cause was presented and approved by user (Step 2 gate)
• For Moderate+ fixes: plan.md exists in .sage/work/
• Test output is PASTED in this response (not summarized)
• Fix is contained to planned files (no scope creep) If ANY fails → go back. Do NOT present the checkpoint.

Sage: Fix verified.

• Root cause: [what was wrong]
• Scope: [Surgical/Moderate/Systemic]
• Change: [what was changed, in which files]
• Tests: [X passed, 0 failed — from actual output] Decision: [root cause + fix approach]. (prepend to the initiative's decisions.md)

[A] Approve — commit and close; branch stays unmerged [M] Merge to [default] — user-gated merge per git-discipline [R] Revise — something's not right [V] Verify — type /review for independent check

Pick A/M/R/V, or tell me what to change.

[M] is the ONLY merge path — [A] never merges. On [M], apply the merge protocol from sage/core/capabilities/execution/git-discipline/ SKILL.md — preconditions, the merge, conflict handling, and the deletion offer all live there; do not restate them here. Omit [M] when not a git repository or no initiative branch exists.

On approval — Post-Flight (Rule 7):

1. Prepend root cause and fix to the initiative's decisions.md (cross-initiative decisions go to the global .sage/decisions.md)
2. Update artifact frontmatter if relevant
3. Store root cause and fix in memory (tagged self-learning) with WHEN/CHECK/BECAUSE prevention rule
4. Wiring check: Verify the fix is fully connected — if you changed an interface, check all callers. If you added error handling, verify the error path is wired end-to-end. Incomplete wiring is the #1 cause of fixes that need re-fixing.
5. Ontology update (if sage-memory available): If the fix changed module dependencies or interfaces (e.g., a service now calls a different service, a dependency was added/removed), update ontology relations. Most Surgical fixes won't need this — only update when the codebase's structural relationships changed.
6. Next steps (Zone 3):

Next steps: /reflect — review what went wrong, prevent recurrence /review — independent evaluation of the fix /build — spec → plan → implement (if fix revealed a feature need)

Type a command, or describe what you want to do next.

Quality Criteria

Communication style: Diagnostic precision. State root cause clearly, explain the chain of causation, describe the fix scope assessment, and verify with evidence.

Good fix output:

• Root cause identified with evidence, not just symptoms addressed
• Fix scope classified honestly (not downgraded to skip the plan)
• A reproducing test exists before the fix is applied
• Verification output is pasted (not summarized)
• If the bug has related patterns elsewhere, those are flagged
• The fix is minimal — no unrelated refactoring mixed in

Rules

• Root cause before fix (Step 2 gate). DO NOT fix before root cause is confirmed with evidence.
• Scope before implementing (Step 3). Classify impact honestly.
• For Moderate+ fixes: plan.md MUST EXIST before implementation. "I know what to change" is NOT a plan file.
• Tests before code (Base Principle 1). Write failing test first.
• Verify with evidence (Rule 5). PASTE actual test output.
• Capture corrections (Rule 6). Store as self-learning.
• Minimal change: fix the bug, don't refactor the neighborhood.
• If stuck, use problem-solving skill. Don't retry the same approach.
• If the fix reveals a systemic issue, ESCALATE. Do not scope-creep a fix into a rebuild. Offer /build or /architect.

Failure Modes

• Agent skips root cause investigation: Red flags section catches this pattern. The ROOT CAUSE GATE blocks proceeding.
• Agent underestimates fix scope: Escalation signals list catches the most common patterns. Anti-downgrade language blocks "I'll just quickly change these 5 files."
• Fix grows during implementation: Scope guard in Step 4 detects when the fix expands beyond the plan and forces a decision.
• Agent claims tests pass without running them: Self-check requires PASTED output. No paste = checkpoint not ready.