Skip to content

Cross-run workspace substrate + reference agent - #26

Merged
nyashkn merged 5 commits into
mainfrom
claude/session-03914w
Jul 22, 2026
Merged

nyashkn merged 5 commits into
mainfrom
claude/session-03914w

Conversation

@nyashkn

@nyashkn nyashkn commented Jul 22, 2026

Copy link
Copy Markdown
Owner

What

Adds the cross-run workspace substrate — the layer that lets a memoryless
agent remember across runs — and a reference agent that demonstrates the
full read-history → diff → report loop over the seeded growth_demo fixture.

A single run bundle can't answer "when did this claim_id last appear", "how
has this number drifted", or "was this claim refuted in a prior run". The
substrate answers all three, and closes a confirmed hole: a critic-refuted
claim could silently republish in a later run because the guard was in-memory
per-run only.

New surface (insight_kit.platform.gate)

  • new_run_dir — dated, collision-safe run isolation under a workspace.
  • seal_run / list_runs / reindex_runs — an append-only, regenerable
    runs.jsonl manifest (V7 analog: rebuildable from the run bundles).
  • claim_history / claim_by_id — query a claim_id's sightings across sealed
    runs in canonical (completed_at, run_id) order.
  • standing_refutations — claim_ids whose latest verdicted sighting is
    refuted (an unverdicted re-emission does not clear a refutation — that's
    the republish hole).
  • guard_republished_claims — the persistent guard: for each draft/published
    claim carrying a standing refutation, it emits a critic refuting the new
    record and applies a high-severity critique (V16 record-then-enforce;
    published tier → downgrade_required=True). It surfaces, never blocks —
    doctrine holds: only deductive identities block.
  • Typed errors RunNotSealedError, WorkspaceNotFoundError (never guess — an
    absent fact is not a reconstructable one).

Reference agent (examples/growth_demo/reference_agent.py)

Two "pulse days" over the fixture, holding nothing in memory between them:

[pulse-2026-01-30] 30d window: blended_cac=1,207  mer=6.76  naive_cac=1,059 (refuted)
[pulse-2026-02-14] 45d window: blended_cac=1,294  mer=6.63  naive_cac=1,112 (refuted)

blended CAC drift day1->day2: +88 KES (+7.3%)
guard: republished-after-refutation caught for DEMO-D-012

Day 2 reconstructs day 1 entirely from the sealed workspace — the pattern an
external, ephemeral pulse agent would use. Run it with
python -m insight_kit.examples.growth_demo.reference_agent.

How it was built

The substrate was implemented, then put through an adversarial find→verify
review that reproduced 17 defects (0 false positives); all are fixed in
the second commit (guard idempotency + critic-id collision space, V16 budget
preservation, board-audience gating, dot-only run_id rejection, symlink-safe
seal, canonical ordering independent of file order, never-trust-a-stale-index,
typed workspace errors). Then a /simplify pass (asdict serialization, one
manifest-write helper, cheap idempotency scan, single sort key) and a
consolidation of the duplicated claim-id minter into one libs.validation
helper shared by the adapter and the guard.

No changes to the canonical immutable-record store; the workspace layer is
purely a regenerable index above it.

Tests

916 passed, 1 deselected. New coverage: 49 workspace tests (guard idempotency,
multi-hit, guard→seal→next-run end-to-end, budget preservation, board
audience, at-cap branch, symlinked runs, out-of-order reindex, interrupted-seal
recovery, deleted manifest, nonexistent workspace) + 2 reference-agent tests
(drift read-back, guard catch asserted on the guard's own findings, fixed-
timestamp reproducibility).

🤖 Generated with Claude Code

https://claude.ai/code/session_01Joez4MiYjd6owPcRQNcKrv


Generated by Claude Code

claude added 5 commits July 22, 2026 19:25
…epublish guard

Adds the workspace layer above the single-run gate: dated run dirs under
workspace_dir/runs/, an append-only regenerable runs.jsonl manifest
(seal_run / list_runs / reindex_runs, V7 analog), claim_history and
claim_by_id queries across sealed runs, and standing_refutations +
guard_republished_claims — the persistent guard that surfaces a
critic-refuted claim_id being republished in a later run (emits a
refuting critic claim with prior-run provenance and applies a
severity=high critique, V16 record-then-enforce; surfaces, never
blocks). Closes the deferred review finding that the refutation guard
was in-memory per-run only.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Joez4MiYjd6owPcRQNcKrv
…ish guard

An adversarial find->verify review of 5282101 surfaced 17 reproduced
defects (0 refuted), collapsing to 8 fixes:

- guard was not idempotent and could crash: a re-invocation or a
  900-bucket critic-id hash collision re-emitted a duplicate claim_id and
  raised LayerAValidationError out of a surface-never-block function.
  Critic ids are now 9-digit and salted with the target record id; the
  guard reuses an already-emitted critic instead of re-emitting, and wraps
  emit in try/except as a safety net (critic_record_id=None branch).
- guard spent the shared V16 critiqueRounds budget, so draft-tier
  surfacings could silently downgrade a later legitimate high critique;
  the counter is now saved/restored around the guard's apply_critique.
- guard hardcoded audience=None, letting board-audience drafts escape the
  gate; it now passes the record's real audience.
- new_run_dir accepted dot-only run_ids ("..") that wrote bundles into the
  workspace root; rejected now.
- seal_run could not seal a run under a symlinked runs/ dir (resolve()
  asymmetry); the runs root is now resolved on both sides.
- "latest" sighting depended on manifest file order, so a reindex could
  flip standing_refutations; order is now canonical (completed_at, run_id).
- _build_entry silently treated a missing/stale records.jsonl as no
  claims, poisoning the guard; it now regenerates via store.reindex (V7)
  and raises on genuine corruption.
- a typo'd workspace path silently answered "no refutations"; queries now
  raise WorkspaceNotFoundError, and a deleted-but-regenerable manifest
  falls back to an in-memory bundle scan without rewriting on read.

Adds 15 regression tests (guard idempotency, multi-hit, guard->seal->next
run end-to-end, budget preservation, board audience, at-cap branch,
symlinked runs/, out-of-order reindex, interrupted-seal recovery, deleted
manifest, nonexistent workspace, fixed-space critic id). Full suite 914
passed.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Joez4MiYjd6owPcRQNcKrv
…nstruction

Quality-only cleanup from a /simplify pass (no behavior change; 914 tests
still green):

- serialize rows via dataclasses.asdict instead of hand-written
  _sighting_to_dict/_entry_to_dict mirrors; the deserialize side stays
  explicit for its coercion/defaulting.
- one _write_manifest_rows(mode=) helper for both the seal-append and
  reindex-overwrite sites, so the V7 'seal row == reindex row' invariant
  cannot drift between two copies.
- seal_run idempotency check scans run_id off the raw manifest lines
  instead of deserializing every prior run's full claim list.
- single canonical (completed_at, run_id) sort key, defined once and used
  by both _scan_bundles and list_runs.
- build RepublishFinding once after the reuse/emit branches rather than
  from two near-identical literals.
- honest docstrings on _build_entry/RunEntry: records.jsonl order is
  emission order on a fresh bundle but record_id order after an
  index-repair reindex (verdict resolution is order-independent).

Deferred cross-module extractions noted for follow-up (shared claim-id
minter in libs/validation; paired read_run_json + shared jsonl helpers in
store/runstate; splitting apply_critique's record-vs-spend concerns).

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Joez4MiYjd6owPcRQNcKrv
Consolidate the duplicated 'namespace + tier token + sha256 bounded number'
id grammar into one libs.validation.mint_claim_id(namespace, tier_token,
seed, *, digits) — the /simplify altitude + reuse pass flagged two divergent
copies (hamilton adapter _gen_metric_claim_id at 3 digits; workspace
_guard_critic_claim_id at 9) that would drift from CLAIM_ID_REGEX and each
other over time.

- adapter._gen_metric_claim_id now delegates (digits=3); output is
  arithmetically identical to the old inline form, so generated metric ids
  are unchanged. Drops the now-dead _ID_TIERS classvar and unused re import.
- workspace._guard_critic_claim_id delegates (digits=9), keeping its
  namespace extraction and wide collision-resistant space; drops the unused
  hashlib import.
- the id grammar's tier-token set now lives once beside CLAIM_ID_REGEX.

No behavior change to persisted ids; 914 tests green.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Joez4MiYjd6owPcRQNcKrv
…loop

A worked example over the growth_demo fixture showing what the cross-run
substrate is for: a memoryless agent that reconstructs context from sealed
runs. Two 'pulse days' — day 1 runs the DAG, emits CAC/MER/naive-CAC claims,
reconciles each against the fixture ground truth (true CAC supported, the
P5-trap naive CAC refuted), seals; day 2 re-runs on a longer window, reads
the blended-CAC drift back from claim_history, and runs the persistent
republish guard, which catches that the naive claim — refuted on day 1 —
was re-emitted without a superseding verdict.

Nothing is held in memory between runs: day 2 reads day 1 back from the
sealed workspace (claim_history / the guard), mirroring how an external,
ephemeral pulse agent would work. Ships with tests pinning the drift
read-back, the guard catch (asserted on the guard's own findings, not the
reconciliation edge), and reproducibility under a fixed emit_timestamp.

916 tests green.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Joez4MiYjd6owPcRQNcKrv
@nyashkn
nyashkn merged commit a1583e8 into main Jul 22, 2026
6 checks passed
@nyashkn
nyashkn deleted the claude/session-03914w branch July 22, 2026 22:45
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

2 participants