investigate

name

investigate

description

The Infiquetra systematic-debugging engine. Find the root cause of a bug, test failure, error, or unexpected behavior before any fix. Diagnosis is the primary deliverable: it writes an agent-consumable DEBUG REPORT, routes real implementation work to saga:work via a saga:handoff issue, design problems to saga:brainstorm, and trackable defects to saga:handoff. Read-only on the world and the saga; never commits, pushes, opens a PR, or deploys.

Investigate

/investigate answers "What is the ROOT CAUSE — and is there a fix small enough to apply safely here?" It investigates a bug systematically, traces the full causal chain from trigger to symptom before proposing anything, and treats the diagnosis as the deliverable. The fix is optional, gated, and deliberately small; real implementation work routes to /work.

This engine is a faithful 3-source merge. The spine — investigate-before-fixing, the causal-chain gate, and predictions-for-uncertain-links — is a PORT of CE ce-debug. The pattern-signature table and the enum'd DEBUG REPORT status (DONE / DONE_WITH_CONCERNS / BLOCKED) are GRAFTS from gstack investigate. The Iron Law ("no fixes without root-cause investigation first") and the multi-component boundary instrumentation technique are BORROWED from superpowers systematic-debugging (self-contained here — superpowers is not a runtime dependency).

Position in the lifecycle

/investigate is OFF-CHAIN (like /loop, /doc-review, and /founder-review): it is not a lifecycle phase, it does not advance lifecycle_phase, and it writes no saga tick. It is a routed diagnostic lens reachable three ways:

• (a) directly — /investigate <error | test | issue# | description>;
• (b) via /loop — when the router meets a failure that wants root-cause work;
• (c) via /qa — /qa routes deep post-merge failures here for root-cause analysis (/qa's Phase 6.2 routes deep post-merge root-cause failures here).

It is not on the linear spine (loop/references/dispatch-table.md). Its outputs route back onto the spine: an applied trivial fix ships through /work or /code-review; a real fix becomes a /handoff issue that /work then executes; a design problem goes to /brainstorm.

Core principles

1. Investigate before fixing — the IRON LAW. NO FIX until you can explain the full causal chain from trigger to symptom with no gaps. "Somehow X leads to Y" is a gap. The causal-chain GATE (Phase 2) is the enforcement: you do not enter Phase 3 until the chain is complete. Fixing a symptom creates whack-a-mole debugging; every symptom patch makes the next bug harder to find.
2. Predictions for uncertain links. For each uncertain or non-obvious link in the chain, form a prediction — something in a different code path or scenario that must also be true if the link is correct. A wrong prediction with a fix that "works" means you found a symptom, not the cause — the real cause is still active. When the chain is obvious (missing import, explicit null deref), the chain explanation itself is the gate; a prediction is for uncertain links only, never a ritual.
3. Evidence before hypothesis. Run an assumption audit first — list every "this must be true" belief and mark it verified (you read the code, checked state, or ran it) or assumed; assumptions are the most common source of stuck debugging. Then instrument the boundaries: capture the actual values at each component boundary, never assumed ones. "X seems off" is not evidence; "X equals null at file:42 because Y was never initialized under condition Z" is.
4. One change at a time; when stuck, diagnose why. No shotgun debugging — one hypothesis, one change, one test. When stuck, do not try harder; diagnose why you are stuck. Two distinct numeric gates govern this (kept separate — see below): the Phase 2 hypothesis-exhaustion gate and the Phase 3 3-failed-fix-attempts gate.
5. Diagnosis-PRIMARY; the fix is GATED, trivial-scope, and self-verified. The DEBUG REPORT is the deliverable. The user chooses fix-vs-diagnosis. Only a trivial / single-concern fix is applied inline (test-first + the engine's own verification); any real implementation work routes to /work. /investigate never commits, pushes, opens a PR, or deploys.
6. Read-only on the world and the saga. gh and saga.py are used read-only (restore / ticks for evidence; never mint, never save). Real fixes route to /work; trackable defects route to /handoff; design problems route to /brainstorm. Large or parallel read-only investigation is offered a backend (operator-choice), never auto-launched.
7. Leave the system smarter. The deliverable is an agent-consumable DEBUG REPORT a future agent can act on cold. A non-obvious root cause is promoted (gated, selective) to the engineering journal LEARNINGS.md; a localized mechanical fix with no generalizable insight is skipped silently.

The two distinct numeric gates

These are two separate counters — do not merge them into one "3-strike" rule:

• Phase 2 — HYPOTHESIS-EXHAUSTION gate. When 2-3 hypotheses are exhausted without confirmation, STOP and run the smart-escalation table (references/methodology.md). This is the gstack "if 3 hypotheses fail → STOP, this may be architectural" numeric form. It counts hypotheses.
• Phase 3 — 3-FAILED-FIX-ATTEMPTS gate. After 3 applied fixes fail, question the architecture and return to root-cause (Phase 2). This counts applied fix attempts, not hypotheses. A failed fix with a wrong root cause is a Phase-2 problem wearing a Phase-3 costume.

Interaction method

Follow ../../references/operator-choice.md for choices from a known set (fix-vs-diagnosis-vs-rethink, an execution backend for large/parallel read-only investigation, the trivial-bug fast-path apply/skip). Ask one question per turn; never silently skip a question. Do not ask questions by default—investigate first (read code, run tests, trace). Ask only when a genuine ambiguity blocks investigation and cannot be resolved by reading code or running tests.

In a channel session (redis-channel active), inline the choices in your reply text instead, following the canonical channel-inline convention in saga/skills/brainstorm/SKILL.md (do not duplicate its wording here).

Use repo-relative paths in every generated document. Absolute paths break portability across machines and worktrees. (The one exception is a saga --saga-id value — a derived id, not a path.)

---

Phase 0 — Triage

Parse the input and reach a clear problem statement. The input is an error message, a test path, a GitHub issue ref, or a free-text description of broken behavior. Take it from command arguments or the active artifact.

Issue reference (#N, org/repo#N, a github.com URL): fetch it read-only and read the full comment thread, not just the opening post:

gh issue view <N> --json title,body,comments,labels

Comments frequently carry updated repro steps, narrowed scope, prior failed attempts, or a pivot to a different suspected cause; treating the opening post as the whole picture sends the investigation the wrong way. Everything else (stack traces, test paths, error messages, descriptions) is the problem statement.

Trivial-bug fast-path. If the cause is immediately readable from the input (single-file typo, missing import, obvious null deref / off-by-one with a one-line fix) and verification needs no deep tracing, present the cause and the proposed one-line fix, then run the Phase 2 fix-vs-diagnosis gate before editing — the fast-path saves investigation ceremony, not the user's choice. If the user picks fix, run Phase 3's workspace/branch check, apply it, self-verify (fresh-reproduce), and skip to the Phase 4 DEBUG REPORT. If diagnosis-only, write the report and stop. When in doubt, run the full framework — a wrong root cause costs more than the ceremony.

Prior-attempt awareness. Only when the user signals prior failure ("keeps failing", "I've been trying", "stuck"), ask what they already tried before investigating — this avoids repeating dead ends and is one of the few cases where asking first is right.

Saga evidence (READ-ONLY). If a work-thread saga exists for this change, read it for evidence — never mint or write it:

python3 plugins/saga/scripts/saga.py restore --saga-id <issue-N|task-slug>
python3 plugins/saga/scripts/saga.py ticks --saga-id <issue-N|task-slug>

Then proceed to Phase 1.

---

Phase 1 — Investigate (read-only on the world)

See references/methodology.md for the full recipes.

1.1 Reproduce. Confirm the bug exists and understand its behavior — run the test, trigger the error, follow the reported steps. For a UI surface, drive the running app with the installed chrome-devtools / playwright MCP (read console + network). If it does not reproduce after 2-3 attempts, use the intermittent-bug techniques in the reference; if it cannot reproduce at all here, document what was tried and what conditions appear missing.

1.2 Environment sanity. Before deep tracing, confirm the environment is what you think — branch, dependencies installed/up-to-date, expected runtime version, required env vars present, no stale build artifacts, dependent local services running when the bug plausibly involves them. Stale node_modules / artifacts are a frequent false lead. The full checklist is in references/methodology.md.

1.3 Trace backward. Trace data flow backward from the symptom to where valid state first became invalid. Read the stack trace bottom-to-top (the bottom frame is the symptom; the cause is upstream), then instrument the boundaries around the first frame with invalid data to capture the actual values. For a bug crossing subsystems, use multi-component boundary instrumentation — log what enters and exits every boundary in one run and read the log linearly; the boundary where data first stops matching expectation is the failing layer. Do not stop at the first function that looks wrong — the cause is where bad state originates.

1.4 Recent changes + evidence. As you trace, check recent changes (git log --oneline -10 -- <file>, git blame). If the bug is a regression ("it worked before"), use git bisect (recipe in the reference). Harvest evidence from gh (issue/PR/check output, read-only), application logs, the saga ticks from Phase 0, and any prior-session skeletons — all read-only.

---

Phase 2 — Root cause

Reminder: investigate before fixing. No fix until the full causal chain has no gaps.

Assumption audit (first). List every "this must be true" belief your understanding depends on and mark each verified or assumed (principle 3). Many "wrong hypotheses" are correct hypotheses tested against a wrong assumption.

Pattern-signature match. Check the symptom against references/pattern-taxonomy.md — gstack's 6 signatures (race / null / state-corruption / integration / config-drift / stale-cache) plus the infiquetra serverless row (cold-start / IAM-permission / eventual-consistency-DynamoDB / throttling / env-or-SSM-drift). Pattern-matching is cheap and points the trace at the right place first.

Form hypotheses, ranked by likelihood. For each:

• what is wrong and where (file:line);
• ≥1 concrete observation that supports it — a runtime value, a log line, an instrumented boundary capture, a behavior delta against a working case. "X seems off" is theorizing; go back to Phase 1 and instrument;
• the step-by-step causal chain from trigger to symptom;
• for uncertain links — a prediction (principle 2). Obvious chains skip the prediction.

Causal-chain GATE. Do not proceed to Phase 3 until the full chain — trigger through every step to symptom — has no gaps. The user may explicitly authorize proceeding with the best-available hypothesis if investigation is stuck.

Hypothesis-exhaustion gate (Phase 2's numeric gate). When 2-3 hypotheses are exhausted, STOP and run the smart-escalation table (references/methodology.md): different-subsystem scatter → design problem (suggest /brainstorm); self-contradicting evidence → wrong mental model, re-read without assumptions; works-locally-fails-elsewhere → environment problem; fix-works-but-prediction-wrong → symptom, keep investigating.

Parallel read-only sub-agents (offer). When hypotheses are evidence-bottlenecked across clearly independent subsystems, OFFER a backend per ../../references/operator-choice.md (inline / verified-workflow) to dispatch read-only probes in parallel — each with one explicit hypothesis and a structured evidence-return format, no edits. Never auto-spawn; skip when hypotheses depend on each other. When authorized, parallel read-only probes use explorer agents. This plugin has no agents/ directory.

Present findings + the gate. Present the root cause (causal-chain summary with file:line), the proposed fix and which files would change, the recommended regression test, and whether existing tests should have caught it. Then run the fix-vs-diagnosis-vs-rethink gate through the native operator-choice contract:

1. Fix it now — proceed to Phase 3 (only applies the fix if it is trivial/single-concern; else diagnose + route to /work).
2. Diagnosis only — I'll take it from here — skip Phase 3, write the report (Phase 4), stop.
3. Rethink the design — only when the root cause reveals a design problem (a wrong responsibility or interface, wrong/incomplete requirements, or every fix is a workaround). Route to /brainstorm. Size alone never makes a bug a design problem.

---

Phase 3 — Fix (GATED, optional, trivial-scope, own verification)

Run this phase only if the user chose "Fix it now" and the fix is trivial / single-concern. If the proper fix is real implementation work, diagnose and route to /work instead — do not grind a big change in here.

Reminder: one change at a time.

Workspace / branch check. Before editing: git status — if the user has uncommitted work in files that need changing, confirm before editing. If on the default branch, ask whether to create a feature branch (detect the default via main / master / git rev-parse --abbrev-ref origin/HEAD with the origin/ prefix stripped).

Test-first. (1) Write a failing regression test that captures the bug; (2) verify it fails for the RIGHT reason — the root cause, not unrelated setup; (3) implement the minimal, root-cause-only fix — no drive-by refactors, formatting, or unrelated cleanup; (4) the test passes; (5) run the full suite for regressions; (6) self-review the diff.

Own minimal verification. Fresh-reproduce the original bug and confirm it is fixed — this is the engine's own verification, not optional. /investigate does NOT route to /qa to verify; the acceptance gate is a separate, downstream concern.

On a failed fix. Return to Phase 2 and explicitly INVALIDATE the current hypothesis first — state what evidence ruled it out — then form a new one with its own grounding observation and prediction. Do not retry variants of the same theory (the rationalization spiral).

3-failed-fix-attempts gate (Phase 3's numeric gate, separate from Phase 2's). After 3 applied fixes fail, the root-cause identification was likely wrong: question the architecture and return to Phase 2.

Blast-radius FLAG. A fix touching >5 files is a FLAG — surface the blast radius to the user and offer a backend — not the inline-vs-route discriminator. The discriminator is trivial/single-concern vs real implementation work; a 6-file mechanical rename can still be trivial, and a 1-file change can still be real work that belongs in /work.

Conditional defense-in-depth. When the root-cause pattern exists in 3+ other files (grep the signature) or the bug would have been catastrophic in production, add guards at the appropriate layers (entry validation / invariant check / environment guard / diagnostic breadcrumb) per references/methodology.md. Skip for a one-off with no realistic recurrence path.

/investigate never commits, pushes, opens a PR, or deploys — even after a successful self-verified fix. Shipping is /work / /code-review's job.

---

Phase 4 — Report + route

4.1 Emit the DEBUG REPORT. Write the agent-consumable report to:

docs/investigations/<slug>-<date>.md

Use the enum'd shape in references/debug-report.md: Symptom / Root cause (causal chain with file:line) / Fix (or "diagnosis only") / Evidence / Regression-test path / Related / Status (DONE / DONE_WITH_CONCERNS / BLOCKED).

4.2 Learning capture — BOTH-SPLIT. Two distinct sinks, by what was found:

• Non-obvious root cause → journal LEARNINGS.md (gated, selective): append a new learning when the lesson is generalizable (a wrong assumption about a shared dependency, a pattern in 3+ locations). Skip silently when the fix is mechanical with no insight. Promotion is a pure append per the journal rules; offer it neutrally when the lesson is one sentence.
• Trackable confirmed defect → /handoff (as a defect-type SDLC issue, --issue-type defect): the report is agent-consumable EVIDENCE, not a handoff source. Open the defect by describing the bug with the report LINKED as evidence inside the issue. NEVER pass the report path to handoff_envelope's classifier — infer_maturity keys on docs/plans/ / docs/brainstorms/ etc. and does not recognize docs/investigations/, so it would fall through to requirements-ready and mis-classify the report. See references/debug-report.md.

4.3 Route per loop/references/dispatch-table.md (read it; never restate it):

• inline fix applied + self-verified → /work or /code-review to SHIP it via a PR (the fix is on a branch; /investigate does not push it);
• real fix needed → /handoff → SDLC issue → /work executes it (NOT by handing /work a docs/investigations/ path — /work consumes a plan path, a GitHub issue, or a resume request);
• design problem → /brainstorm;
• diagnosis-only → stop.

There is NO saga write — /investigate is off-chain and saga READ-ONLY.

---

Hard boundary

/investigate investigates, diagnoses, optionally applies a trivial self-verified fix, writes the DEBUG REPORT, and routes — then stops. It does NOT:

• mutate the world or the saga — gh / git / saga.py are READ-ONLY (restore / ticks for evidence only; never save, never mint);
• commit, push, open / update / merge a PR, or deploy — even after a successful fix;
• file SDLC issues — mission-control owns the SDLC; defects route through /handoff;
• route to /qa to verify — it does its own fresh-reproduce verification; the acceptance gate is downstream and separate;
• write the saga — off-chain, no lifecycle_phase advance, no tick;
• do real implementation work inline — that routes to /work; only trivial/single-concern fixes apply here;
• path-classify the DEBUG REPORT — never pass docs/investigations/... to handoff_envelope; link it as evidence in a described DEFECT instead.

fix is this engine's identity verb — gated-allowed, trivial-scope, self-verified. It is never the default action; the diagnosis is.

---

Reference files

• references/methodology.md — the full procedure: triage (trivial fast-path + prior-attempt); investigate (env-sanity checklist + backward-trace recipe + multi-component instrumentation + git bisect); root-cause (assumption audit + hypotheses-with-evidence + predictions + causal-chain gate + smart-escalation table + the hypothesis-exhaustion gate + parallel read-only sub-agent dispatch via ../../../references/operator-choice.md); fix (workspace/branch check + test-first + minimal-diff + own minimal verification + failed-fix invalidation + the 3-failed-fix-attempts gate + the >5-file blast-radius FLAG + conditional defense-in-depth).
• references/pattern-taxonomy.md — gstack's 6 bug signatures + the infiquetra serverless addition with a symptom→suspect map, and the anti-patterns catalog (shotgun debugging, confirmation bias, "it works now move on", prediction quality bad-vs-good, rationalization-spiral red flags).
• references/debug-report.md — the DEBUG REPORT template (the gstack enum'd shape), the journal- LEARNINGS promotion template, and the /handoff defect-routing note (report = evidence, fix reaches /work via a /handoff ISSUE, never via the report path through the classifier).
• ../../references/operator-choice.md — the 3-backend contract for offering a backend for large/parallel read-only investigation.
• loop/references/dispatch-table.md — the outbound routing reference (read, never restate).
• ../brainstorm/SKILL.md — the canonical channel-inline convention (cite, never duplicate).
• ../../references/saga-spec.md — the saga contract (restore / ticks; /investigate is read-only).