Files

Tom Boucher fa41bfec5c enhance(#3942 ): the emitted-drift ack is PR-lifetime data — move it to a commit trailer (#3954 )

* test(#3942): failing-first suite for the emitted-drift ack commit trailer

Binds 37 input classes from the phase test matrix to the behavior ADR-3942
specifies, before any of it exists. Stubs return benign empty values rather
than throwing, deliberately: several rows assert that something DOES throw
(cap overflow, uncomputable commit range), and a throwing stub would turn
those green for the wrong reason and destroy the red.

The two rows that carry the design's load:

- merge-base semantics. The range is $(git merge-base base HEAD)..HEAD, not
  base..HEAD, because changedPaths comes from `git diff base...HEAD` (three
  dot). Two-dot would let the ack set and the change set disagree about which
  commits are this PR's. The fixture forks a topic branch, puts a trailer on
  each side, and asserts only the topic-side trailer is in range.

- fail-closed on an uncomputable range. With fragments a depth-1 checkout
  passes VACUOUSLY, every fragment reading as brand-new. With trailers the
  range cannot be computed at all, and returning an empty set would silently
  disarm the gate, so it must throw. The fixture builds a genuine shallow
  clone rather than simulating one.

Also covers the self-inflicted case: this change's own documentation quotes
the trailer syntax, so an example landing at the end of a commit message would
arm a live acknowledgment keyed on the literal placeholder text. Keys carrying
angle brackets or whitespace are rejected.

Authored per the phase artifacts 40-design.md and 50-test-matrix.md.
Not yet run on the remote runner — this commit exists to be tested.

Refs #3942

* chore(#3942): move the emitted-drift ack to a commit trailer

Implements ADR-3942, superseding ADR-2719 section 3 and its #2789 amendment.
Sections 1, 2 and 4-7 are retained: the conservation law is unchanged, only the
storage of its escape hatch moved off the working tree.

An acknowledgment explains one PR's ripple, and the moment that PR merges the
ripple is in the base, so it can never clear anything again. It was stored in
permanent shared state anyway, and every consequence of that mismatch had to be
built and then maintained. The chain is #2789 -> #2914 -> #3078 -> #3842 ->
#3823 -> #3875, each fix generating the next defect, ending in a scheduled
sweeper whose own first PR could not merge itself.

Added
  parseAckTrailers + renderAckTrailer (pure) and readAckTrailers (IO shell),
  reading Emitted-Drift-Ack-Hash: / Emitted-Drift-Ack-Growth: trailers over
  the merge-base range. tests/emitted-ack-trailer.test.cjs, 37 cases, written
  failing-first and confirmed red before any of this existed.

Changed
  diffEmitted takes two structurally distinct key-space maps instead of one
  shared paths map. That closes a latent defect: the spaces were separated by
  convention only, so a growth key satisfied a hash lookup by naming
  coincidence. staleAcks now reports which space a key was declared in.
  REMEDIATION teaches the trailer, per space, with its example rendered through
  renderAckTrailer so the taught grammar cannot drift from what the parser
  accepts.

Removed
  the sweep workflow, the guard-no-ack-on-next job, the standalone linter and
  its lint:ci entry, the fragment directory and its three spent fragments, the
  legacy single-file union, and the baseAck/spentAcks mechanism -- spentness is
  now structural, not computed.

Two range properties carry the design and are pinned by tests rather than
asserted: the range is merge-base scoped, matching git diff base...HEAD, so an
already-merged trailer is out of range by construction; and an uncomputable
range throws instead of reading as zero acknowledgments, which is the inverse
of the fragment guard's vacuous pass.

Three deliberate observable changes, each disclosed in the changeset: the
unread runtime field is gone, the legacy file is no longer read, and cross-space
excusal no longer works.

Ten open PRs carry fragments and will meet a modify/delete conflict. Measured
before landing and accepted deliberately; the one-line migration is in the PR
body.

Verified: lint:ci exit 0. Remote runner to follow on this exact sha.

Refs #3942

* fix(#3942): silent trailer collapse, lost coverage, and an unbounded cap

Six findings from the orthogonal review round, all fixed in place.

BLOCKER -- two trailers of the same name on one commit collapsed silently.
readAckTrailers built `separator=1d` where git needs `separator=%x1d`: the
`separator=` value inside a %(trailers:...) placeholder is itself a
pretty-format string, so the bare hex was emitted as two literal characters
and the split on \x1d never matched. Two same-name trailers therefore joined
into one value with errors empty -- the first reason absorbing the second
entry's key. Silent truncation, the exact class MAX_ACK_TRAILERS throws to
prevent. Confirmed with od -c against real git output before and after.

The failing-first matrix did not catch it because its "both spaces coexist"
row uses Hash plus Growth -- different trailer NAMES -- so the value separator
was never exercised. Two regression tests now cover same-name trailers
directly.

Coverage recovered: normalizeAckReason and INVISIBLE stayed on the live path
via parseAckTrailers but lost every test when the old suite was pruned. Back
under test against the current surface -- all six invisible codepoints
individually, whitespace collapse, trim, CRLF, and two seeded fast-check
properties. Dropping any single codepoint now fails.

MAX_ACK_TRAILERS counted raw trailers before de-duplication, so one trailer
carried forward across rebased commits counted once per commit and could throw
on a legitimate branch. Now counts distinct entries; 100 identical repeats
dedupe to one.

diffEmitted validated baseline, current and changedPaths but not the new
ackHash/ackGrowth, so a bad shape raised an unhandled TypeError instead of an
error verdict -- the same defect shape this file documents for #2778.

Docs: CONTRIBUTING and TESTING-SUITES were rewritten only in their first
sections; the later passages still taught fragments, git rm and the deleted
guard, contradicting the new text directly above them. Finished.

Also extends lint-removed-but-needed to exempt docs/adr and docs/research.
That gate fails on any docs mention of a file deleted in the same diff, which
makes it impossible to document a deletion in the PR performing it -- an ADR's
whole job is naming what it retired. Exemption is narrow and comes with a test
proving the gate still fires for a live consumer elsewhere under docs/. A
guard that cannot fail is worse than no guard. Maintainer-approved.

CONTEXT.md names the retired machinery by role rather than by filename: its
generated projection lands in docs/, which that gate does scan.

Adds docs/how-to/acknowledge-emitted-drift.md. The required docs set is
Reference and Explanation, so the task quadrant can be empty with every gate
green -- and this change has a real multi-step journey, including the fragment
migration ten open PRs now need.

lint:ci exit 0.

Refs #3942

* docs(#3942): correct the duplicate-trailer rule in CONTRIBUTING

Both axes of the code review independently flagged the same passage, without
seeing each other's output.

It claimed two declarations of the same key are always "a hard, loudly-reported
error, not a silent last-wins". That is only half true, and the missing half is
the one contributors hit: identical declarations -- same key, same reason --
dedupe silently, because a trailer legitimately survives a rebase and reappears
on every rebased commit. Failing there would red a branch for doing nothing
wrong, which is exactly why the dedup exists.

Only a same-key/different-reason pair errors, and that one is a genuine
ambiguity about which explanation holds.

As written, the paragraph told a contributor that a rebase-carried trailer
breaks the gate -- the opposite of the behavior. CONTEXT.md's parallel entry
already stated it correctly; this brings CONTRIBUTING into line.

Doc-only, root-level markdown.

Refs #3942

* chore(#3942): backfill changeset PR number to 3954

---------

Co-authored-by: sim <sim@local>

2026-08-27 17:28:39 -04:00

18 KiB

Raw Blame History

ADR-3128: Adaptive runtime evidence for GSD Debug

Status: Proposed
Date: 2026-08-07
Extends: ADR-1671 (dynamic context management platform — this ADR is a coordinated WHEN_VOCABULARY amendment, see Architecture and contracts), ADR-1610 (workflow/agent size budget ratchet)
Relates: ADR-1577 (untrusted-input boundary — captured events are untrusted data), ADR-227 (validate shape, not just type). Issue #3128 is the scope authority; #1957 (debugger hardening) and #1961 (bug-taxonomy routing) are completed foundations, not approval for this scope; #3149 delivered the cmdInitDebug prerequisite.

Context

/gsd:debug routinely reaches several plausible hypotheses that static inspection, the existing test suite, and existing logs cannot distinguish. At that point a solo developer has two bad options: hand-add instrumentation, or let the agent guess. Hand-added instrumentation is easy to forget, contaminates a commit, and is especially unsafe in a dirty worktree. Guessing produces a plausible-but-wrong root cause and an unproven fix.

GSD already has persistent debug sessions, bug-class routing, context-reset recovery, TDD, five-signal fix acceptance, and human verification. What it lacks is a bounded protocol connecting a recorded hypothesis to temporary runtime observations, comparing the identical reproduction before and after a fix, and proving every diagnostic edit and artifact was removed before completion.

Why this needs an ADR rather than an issue alone

Three of the four things being introduced are long-lived architectural commitments, not implementation detail:

A persisted schema (schema_version: 1) written into user projects' .planning/debug/*.md files, which must stay readable by every later version.
An ownership model for agent-authored edits to tracked source — the one operation in this tool that can destroy uncommitted user work.
A cleanup state machine that gates terminal transitions — it can refuse diagnosis completion, human verification, abandonment, archive, staging, commit, and knowledge-base writes.

Additionally, ADR-1671's WHEN_VOCABULARY is closed: every growth from 14 to 29 entries was a coordinated amendment, never an organic edit. Admitting an atom therefore requires an amendment, and this ADR is it.

What already exists, after #3149

init.debug (cmdInitDebug, src/init.cts) — a dedicated entry point, so ADR-1671 admission gate (2) ("a fact the init seam demonstrably computes at a real entry point") is satisfied for debug. Before it, any debug-scoped atom evaluated false forever.
gsd-core/workflows/debug.md Step 0 already consumes one init bundle and already documents the section_manifest read contract, including that null means read everything and is not the same as an empty included array.
gsd-core/workflows/section-manifest.json has no debug key, because debug.md carries no section markers yet. Admission gate (1) — a named consuming section of at least 400 bytes — is what this ADR's implementation must satisfy.

Decision

The applicability atom is a resolved boolean, state:runtime-evidence-eligible — never the raw flag. when= takes exactly one operator-free atom, but policy precedence is explicit flag → valid saved session policy → default. A resumed session (continue <slug>) that already persisted policy: adaptive passes no flag on the resume invocation, so a flag-keyed atom evaluates false on exactly the sessions already running the protocol — ADR-1671's silent-exclusion failure arriving through a different door. Only a resolved boolean can see the saved policy. The fact is folded once in cmdInitDebug as policy !== 'off', following the state:chunked-mode precedent that a compound resolves in the FACT, never in the grammar.
The session gains an immutable goal and an optional Runtime Evidence section at schema_version: 1. Legacy sessions are read without migration-only rewrites: an absent goal means find_and_fix; an absent Runtime Evidence section means off + not_used. A find_root_cause_only session never offers or applies a fix and never edits tracked source.
Every agent-authored source edit is ledgered before it is made, and cleanup is fail-closed. Paired non-nested markers carrying exact gsd-debug-probe:start|end <slug> <probe-id> payloads, plus a pre-edit SHA-256 over the complete raw UTF-8 block including both marker lines and the file's existing line-ending form. Cleanup removes only a complete balanced block whose bytes still match its saved hash; anything else is cleanup_failed, and edits outside the owned block are preserved.
cleanup_failed gates every terminal transition. Diagnosis completion, human verification, abandonment, archive, staging, commit, knowledge-base writes, and all terminal returns are blocked while the ledger is non-clean. The session stays resumable until cleanup is proved. This holds for direct debugger invocation, not only the managed path.
The protocol body is one contiguous fragment; the policy checks stay inline. This is what makes the atom admissible rather than a repeat of flag:--verify-only, which was surveyed and rejected for being interleaved across three non-contiguous touch-points.
The agent-side share of the protocol is bounded by the absence of any lazy-load path. agents/*.md pull references with eager @-includes and the fragment platform is workflow-only, so extraction relocates bytes without reducing per-dispatch cost. Agent files carry routing and gates only; the protocol body lives workflow-side.
Nothing is added that must be operated. No daemon, collector, server, telemetry, network transport, external dependency, hosted service, SDK, or shared application-runtime trace.
The shipped default is off; probes are opt-in per invocation. --runtime-probes selects adaptive; no flag and --no-runtime-probes both select off. The force policy is retired as unreachable and redundant — once probes are opt-in, "the user asked for probes" and "consider probes where safe" are the same intent, and a policy whose meaning is bypass the adaptive safety reasoning should not exist. Every safety precondition still applies to adaptive.

This follows the shipped workflow defaults, which split cleanly by kind: verification gates default on (research, plan_check, verifier, nyquist_validation, security_enforcement), agent autonomy defaults off (auto_advance, research_before_questions, plan_bounce, cross_ai_execution). Installing temporary probes into tracked source without a second confirmation is autonomy, not a gate; cross_ai_execution: false is its closest analogue. Defaulting it on would make it the only capability in that block that edits the user's source unasked. #3128's own Alternative 3 offered the choice — "maintainers can choose a stricter default during approval if desired" — and it is taken here.

Options considered

Keep GSD passive-only. Rejected: static evidence often cannot separate falsifiable hypotheses, so the debugger guesses or delegates instrumentation to the human.
Always install source probes. Rejected: observer effects, concurrency semantics, dirty worktrees, privacy, and cleanup risk make always-on instrumentation unsafe.
Adaptive by default. #3128's proposed posture, and rejected: it would let the debugger edit tracked source on an ordinary invocation nobody opted into, against a config block whose autonomy settings are uniformly off. See Decision 8.
Gate on flag:--runtime-probes. Rejected — see Decision 1. It is the ADR-1671:125 silent-exclusion bug reintroduced through the front door.
Put the protocol body in gsd-core/references/ and @-include it. Rejected on its own terms: an eager include relocates bytes without reducing loaded context, so the size gate passes while the thing it protects gets worse. A reference read on demand at activation time remains correct and is used for the deep protocol detail.
Adopt or vendor millionco/debug-agent. Rejected: its package, installation, runtime and logging contracts add an external dependency and do not fit GSD's self-contained, generated multi-runtime prompt architecture. Concepts are adapted; no code is copied.
Write raw logs into the session or a trace file. Rejected: unbounded, potentially sensitive, and it contaminates durable shared context.

Architecture and contracts

1. The applicability atom

WHEN_VOCABULARY 29 → 30, one entry: state:runtime-evidence-eligible.

Layer	Contract
Grammar	`src/workflow-fragments.cts` — one atom, no operator, no negation, no nesting
Predicate	`src/section-manifest.cts` — `WHEN_PREDICATES['state:runtime-evidence-eligible'] = (facts) => facts.runtimeEvidenceEligible === true`, plus the `InvocationFacts` field
Fact	`src/init.cts` — `cmdInitDebug` folds explicit flag → valid saved session policy → `off` default, then emits `runtime_evidence_eligible = (policy !== 'off')`
Flags	`src/init-command-router.cts` — the `debug:` route's `parseNamedArgs` boolean list gains `runtime-probes` and `no-runtime-probes`; conflicting flags fail closed

A coordinated-change guard in src/section-manifest.cts throws at module load in both directions, so a half-edit cannot ship. gsd-core/workflows/section-manifest.json is regenerated by scripts/gen-section-manifest.cjs --write and verified by --check in lint:ci.

Eligibility is not authorization. The atom decides only whether the protocol text is included. Whether a probe is actually installed on a given run remains subject to every reproduction, privacy, bug-class, perturbation and cleanup precondition, evaluated at run time. force never overrides those, nor diagnose-only mode.

2. Session schema v1

Issue #3128 is the authority for the literal field list. The invariants this ADR locks:

Additive and optional. Absent section ⇒ off + not_used. Absent goal ⇒ find_and_fix. No migration-only rewrites of files already on disk in user projects.
Effective policy precedence is explicit override → valid saved policy → off. An invalid saved value stays on disk for inspection while dispatch fails safe to off — the fail-safe direction is now the same as the default, so an unreadable policy can never widen behavior.
An override changes only policy. It never resets state, mode, probes, artifacts, active_run, or cleanup data. Switching to off still reconciles and cleans an existing non-clean ledger.
Run IDs are allocated write-ahead and never reused. Read next_run_seq: N, persist active_run.run_id: run-N with its phase, exact reproduction reference and start time, advance to N+1, then execute. An interrupted run with no attributable result is appended as inconclusive and cleared before another ID is allocated.
Durable state carries only sanitized facts — counts, hashes, enums, verdicts, references. Raw stdout/stderr, application logs, request/environment data, secrets, credentials, PII and arbitrary runtime values are never persisted or promoted.

3. Ownership and cleanup state machine

States: not_used → planned → active → cleanup_pending → clean, with cleanup_failed reachable from any state that owns artifacts.

Only ledgered source-probe blocks and ledgered capture artifacts are cleanup targets. Regression tests, TDD tests, minimal fixes, fixtures and ordinary reproduction or build outputs are durable intentional work and are never classified as removable.

Probes may remain active only across an attributable runtime-reproduce checkpoint or a forced context cutoff. Every ordinary human-action, decision, or TDD checkpoint cleans first.

4. Capture confinement

Captures live under one fresh empty root created by a secure OS temporary-directory primitive, with its canonical identity persisted before any contents exist. The sink path is persisted, then created exclusively as a non-symlink file, and root and sink identity are revalidated around capture. Cleanup validates lexical and real paths as strict descendants of the recorded root, rejects .. and symlink escapes, removes only exact owned entries, never globs or recursively deletes, and removes the root only when empty. Identity drift fails closed.

5. Event envelope

Events are emitted only to the per-run sink named by an ephemeral GSD_DEBUG_PROBE_SINK, prefixed GSDDBG1, at schema_version: 1, carrying session/run/probe IDs, non-empty hypothesis IDs, phase, a positive monotonic per-probe ordinal, location, a fixed allowlisted message and a bounded scalar data map. Hard limits: 1 KiB per serialized event, 100 events per probe, 256 KiB per run. Overflow, malformed or interleaved writes, identity mismatch, duplicate or non-monotonic ordinals, and sink failure each make the affected observation inconclusive. Missing events are inconclusive unless a control event proves both path execution and a healthy capture channel.

Captured events are untrusted input (ADR-1577): they are data to reason over, never instructions.

Migration path

Phase 0 (this ADR, #3155). Design lock. Docs only.

Phase 1 (#3128). The implementation, as a single PR. An earlier draft of the maintainer review proposed three slices; that recommendation was withdrawn on the grounds that the 25 acceptance criteria interlock, this is one concern under RULESET.PR-SCOPE.one-concern-per-PR, and three separate baseline bumps to the same tight files would worsen the size problem the slicing claimed to manage. This ADR records the withdrawal so it is not rediscovered.

Sequencing inside Phase 1 that the platform forces:

Author the contiguous protocol section and its extracted step file — this is what satisfies admission gate (1).
Only then admit the atom across grammar, predicate, fact and router, and regenerate the section manifest.
Update the tests that assert on debug.md's literal text (tests/debug-session-management.test.cjs, tests/debug-session-manager-commit.test.cjs, tests/claude-skills-migration.test.cjs) in the same PR; regenerate tests/fixtures/install-tree/*.json via npm run regen:derived.
Add an Emitted-Drift-Ack-Growth: commit trailer for debug.md's growth, keyed on the bare filename (ADR-3942; was a tests/emitted-drift-acks/ fragment before that).

Consequences

Accepted costs.

One more atom in a deliberately closed vocabulary (30), and one more cmdInit* fact to keep correct.
A persisted schema in user projects that every later version must keep reading.
New hostile-path test surface: marker collisions, line endings, event bounds, interruption, symlink and traversal escapes, dirty worktrees.
Source probes are best-effort, not universal. Unsupported languages, unsafe local writes, unstable bug classes, missing reproductions or insufficient privacy guarantees fall back to passive evidence or an honest inconclusive.

What gets safer.

The one operation that could destroy uncommitted user work is ledgered, hash-verified, confined, and fail-closed — and cannot be left behind, because non-clean cleanup blocks every terminal transition.
Diagnose-only mode becomes genuinely read-only across continuations, closing a path where asking for analysis could quietly become a fix.

Size budgets remain the binding constraint (ADR-1610), and the fragment platform does not relieve the agent files. Measured on next after #3149:

File	Bytes	Cap	Headroom
`gsd-core/workflows/debug.md`	21,173	40,960	19,787
`agents/gsd-debugger.md`	48,851	57,344	8,493
`agents/gsd-debug-session-manager.md`	19,751	24,576	4,825
`gsd-core/workflows/help/modes/full.md`	38,012	40,960	2,948

Crossing a tier cap means extracting to references/, never a +N bump.

Open questions

Does the maintainer permit the contributor's pre-existing prototype to be used as reference? #3128 discloses a near-complete uncommitted fork-first implementation and makes continuation conditional on an explicit answer. Unanswered as of this ADR; it blocks the contributor, not the design.
~~Should adaptive or off be the shipped default?~~ Resolved: off — see Decision 8. Decided by the maintainer on #3128 after this ADR first merged; recorded here rather than left as an open question, because it is the point an implementer reads first.
How much of the protocol can the two debug agents carry before extraction is forced, given they have no lazy-load path? Answerable only against the real diff, and it is a per-PR review gate either way.

ADR-1671 — the platform this amends; its two admission gates govern the atom.
ADR-1610 — the size ratchet that bounds Phase 1.
ADR-1577 — captured events as untrusted input.
Issue #3128 (scope authority), #3149 (cmdInitDebug prerequisite, merged), #3155 (this ADR's phase).
src/workflow-fragments.cts's section-marker grammar solves several problems this protocol's marker grammar will otherwise rediscover: CRLF-aware per-line splitting for byte-exact reassembly, fence/comment interleaving resolved in one left-to-right pass (a two-pass design silently skips to EOF — DEFECT.CONTEXT-PREDICATES-COMMENT-FENCE-BLIND, #2928), and a frozen REASON enum so tests assert codes rather than message text.

18 KiB Raw Blame History