Make Footwork actions verifiable and recover scoped native appends after restart by Tom-R-Main · Pull Request #1 · Tom-R-Main/Footwork

Tom-R-Main · 2026-10-10T17:31:27Z

Agents can lose an action reply after native input has already been delivered. This change makes the scoped TextEdit append recoverable from a qualified, operation-bound execution receipt, while keeping missing receipts unknown and refusing automatic replay. Exact field content, geometry and resource identity remain checked at the final input boundary.

This is the combined unpublished Footwork development branch. It includes the four-tool capability MCP slice, Jev decision and evaluation infrastructure, bounded browser inspection tools, and the recorded architecture/reference research and experiments. Much of the diff is frozen research evidence; failed and inconclusive runs retain their original conclusions. A merge commit preserves preregistration chronology and source hashes.

The capability slice binds a trusted loopback source and a selected native destination. A detached one-action worker separates execution from the caller transport; private immutable records, write-ahead admission and operation binding let read-only status recover qualified historical completion. Fresh document agreement remains a separate observation. Repeated tokens and status never launch another append.

Independent review and repository bot review reproduced and remediated stale field/save evidence, Unicode receipt corruption, false submission authority from query navigation, and loss of exact field/resource preconditions during live native dispatch. Completion now expires current-state evidence after potentially delivered same-source inputs. Ordinary query navigation supplies no submission acknowledgement, so completion remains blocked where that acknowledgement is required. Browser capture now redacts short opaque URL credentials and retains only documented inventory counters. Frozen visual fixtures are validated against authored hashes across platforms; Mac-only probes remain isolated while portable checks continue on Linux.

Validation:

Full local Python suite after product/evaluation fixes: 1,428 passed, 26 skipped. Subsequent deadline boundary controls: 35 recovery tests passed. Real localhost browser tests: 12 passed.
Independent review of the product changes in one continuing Codex context: 309 focused checks passed. Root checked the reviewed file hashes. Earlier durable-restart design and implementation received Claude Opus 5.5 review through the authorized mailbox; that review predates these PR remediations.
Node tests: 97 passed. Rust formatting, Clippy and 19 tests passed offline. Ruff and notebook validation passed.
Fresh real MCP/TextEdit restart matrix: all 12 cases passed. All three real native interference guards passed: trailing space, trailing newline and changed field geometry.

Evidence: jevdual/results/pr-review-remediation-20261010/. Reference mechanisms and exact local repository revisions: jevdual/docs/plans/reference-restart-reconciliation-20261010.md and reference-capability-build-20261009.md.

The native qualification covers the selected TextEdit AX route and process interruption. It does not establish disk-save durability, power-loss recovery, arbitrary-app support, exactly-once GUI execution or general installation readiness. No provider defaults or deployment are changed by this PR. Footwork remains MIT; dependency licenses continue to apply.

…en_locked, and the posture refuses every delivery; G4 inconclusive (the screen locked mid-run and the harness pressed keys for 25 steps against it) Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_015SXfQdX2QiRjY8NkxiQS5H

…ified-pass column is no longer 0 on every arm (Q17) Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_015SXfQdX2QiRjY8NkxiQS5H

….69, −0.01]) but cost 4 verified passes with no-fact refusals up; the delivery guardrail caught a false not_delivered on select_dropdown; the input fix moved verified passes 121 to 137 Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_015SXfQdX2QiRjY8NkxiQS5H

…s change event confirms a dropdown or input step (Q12b's false not_delivered on every working sort); package D summary Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_015SXfQdX2QiRjY8NkxiQS5H

…ied values, save readback, locked-screen stop) and D's receipt fixes Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_015SXfQdX2QiRjY8NkxiQS5H

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_015SXfQdX2QiRjY8NkxiQS5H

… owner's draft Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_015SXfQdX2QiRjY8NkxiQS5H

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_015SXfQdX2QiRjY8NkxiQS5H

… 6 of 6 (Q17: 0 of 6); the readback polls up to 3 s before calling a save partial Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_015SXfQdX2QiRjY8NkxiQS5H

…re the rerun Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_015SXfQdX2QiRjY8NkxiQS5H

…arly reads; the receipts were true and sent the driver to File > Save; 6 of 6 verified again Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_015SXfQdX2QiRjY8NkxiQS5H

…ifies 148 against 137 with no false completion, but the primary's interval includes zero; not adopted on this run; V1 held Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_015SXfQdX2QiRjY8NkxiQS5H

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_015SXfQdX2QiRjY8NkxiQS5H

…ion rule) and Q11b (the narrow declaration on live-dev) pre-registered Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_015SXfQdX2QiRjY8NkxiQS5H

…w-rust-year on both arms, whose predicate (2015) disagrees with Wikipedia's current infobox (2012) that both arms quoted; the stale oracle is the owner's call Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_015SXfQdX2QiRjY8NkxiQS5H

…ooled through no-fact refusals (91 against 68) with predicate passes unchanged; not adopted Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_015SXfQdX2QiRjY8NkxiQS5H

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_015SXfQdX2QiRjY8NkxiQS5H

…ur source gaps (CLI judge wiring for key and Enter, unbound effect on a raised dispatch, screenshots before scrubbing, strict claim asides); Rust oracle erratum recommended Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_015SXfQdX2QiRjY8NkxiQS5H

… judge through footwork do; a raised dispatch ends in a named refusal; the browser judges Enter like a click; no screenshot while a secret is held; strict claims check facts inside asides Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_015SXfQdX2QiRjY8NkxiQS5H

…arms quoted): Q16b regrades to 20 against 22 of 24 with no false completions, decision unchanged; Q17 corrected to six tasks Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_015SXfQdX2QiRjY8NkxiQS5H

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_015SXfQdX2QiRjY8NkxiQS5H

… facts are elsewhere on the page, before the audit set exists Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_015SXfQdX2QiRjY8NkxiQS5H

…wrong product, wrong field, wrong task, unsupported qualification, fact in an aside, wrong year on the page, order not placed), scored by the claim check and the full verifier in both modes; labels committed before scoring Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_015SXfQdX2QiRjY8NkxiQS5H

…hanged) Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_015SXfQdX2QiRjY8NkxiQS5H

… either mode; strict halves false rejections (6/8 to 3/8); the claim check alone is weak as the audit said (V2 failed); V3 predicts two strict-mode verifier fixes, before making them Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_015SXfQdX2QiRjY8NkxiQS5H

…ies only to tasks that ask for an answer; claims audit false rejection 3/8 to 2/8 with false acceptance 0/10 (V3 failed its threshold; A4's remaining refusal is an absence without a trajectory) Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_015SXfQdX2QiRjY8NkxiQS5H

…lue landed once, at the end, with the earlier content intact, and name the failure otherwise; start clears carries and snapshots; pilot uses both Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_015SXfQdX2QiRjY8NkxiQS5H

…not its value is still exported or long enough to scrub (secondary audit, second round) Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_015SXfQdX2QiRjY8NkxiQS5H

…on, carry, snapshot, receipts, file before and after, recovery), a duplicate variant; G7 before running Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_015SXfQdX2QiRjY8NkxiQS5H

…a traceback; the pilot stops at its first failed step; G7's first run kept as evidence of both defects Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_015SXfQdX2QiRjY8NkxiQS5H

…tests

…dispatch

…ions

…ation

…trials

…ation

chatgpt-codex-connector · 2026-10-10T17:31:38Z

Codex Review Summary

This comment shows the latest Codex review activity on this pull request.

Review	Status	Commit	Review trigger
📝 Code Review	✅ Completed 2026-10-10T17:35:49.209024Z	`678443e`	PR opened

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

Open a pull request for review
Mark a draft as ready
Comment "@codex review" or "@codex security review".

Codex reacts with 👀 while any review is running, comments if it has suggestions, and reacts with 👍 once all reviews finish with no findings.

chatgpt-codex-connector

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 678443eb71

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

Open a pull request for review
Mark a draft as ready
Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

…n CI

…ux qualification

…epair

…speed estimates

Tom-R-Main and others added 30 commits September 26, 2026 15:32

Native results set success from an accepted done, so the report's ver…

a12a43b

…ified-pass column is no longer 0 on every arm (Q17) Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_015SXfQdX2QiRjY8NkxiQS5H

Merge packages E (strict claim check, Q16 arm), F (Q11b arm), G (carr…

91f19f5

…ied values, save readback, locked-screen stop) and D's receipt fixes Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_015SXfQdX2QiRjY8NkxiQS5H

audit_no_fact: read traces through Path (ruff SIM115)

cc0768d

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_015SXfQdX2QiRjY8NkxiQS5H

Legible Loop v6: evidence from packages C to G, rule by rule, for the…

0d3907a

… owner's draft Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_015SXfQdX2QiRjY8NkxiQS5H

Notebook: G5 re-predicts G4 on an unlocked screen, before the run

6918473

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_015SXfQdX2QiRjY8NkxiQS5H

Notebook: G6 predicts G5's first-save partials were early reads, befo…

809e346

…re the rerun Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_015SXfQdX2QiRjY8NkxiQS5H

Merge package E results: Q16 and Q17

a04845e

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_015SXfQdX2QiRjY8NkxiQS5H

Plan status and Legible Loop evidence after Q16b and Q11b

c2de17e

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_015SXfQdX2QiRjY8NkxiQS5H

Q17 write-up: six tasks

8c3009f

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_015SXfQdX2QiRjY8NkxiQS5H

Notebook: V2 predicts the claim check accepts adversarial dones whose…

963aab0

… facts are elsewhere on the page, before the audit set exists Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_015SXfQdX2QiRjY8NkxiQS5H

Claims audit labels: quote task strings (YAML syntax only; no label c…

0af2cae

…hanged) Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_015SXfQdX2QiRjY8NkxiQS5H

Tom-R-Main added 18 commits October 10, 2026 12:14

Record detached feasibility outcome before implementation

3d48261

Amend autosave identity and deadline rules before qualification

2a47557

Record worker and caller deadline ordering before qualification

696a641

Reconcile native appends from durable independent execution receipts

8a407f1

Record invalid fixture placement before fresh restart qualification

320c3b0

Qualify restart recovery with native interruptions and Claude sign-off

c9ba496

Predict combined PR completion and native dispatch remediations

8312cd3

Predict fail-closed submission evidence correction before regression …

537d037

…tests

Predict ordered save evidence and remapped field correction

9b980df

Predict final native resource identity guard before input tests

daa0846

Preserve exact native append and resource preconditions through live …

c0c315d

…dispatch

Keep completion evidence current and literal without inferred submiss…

9961eda

…ions

Preregister native requalification after combined PR fixes

c4d4953

Predict atomic restart harness boundary publication before requalific…

3064be9

…ation

Publish complete restart markers and add real native freshness guard …

4fb9a95

…trials

Predict delayed AX geometry publication before guard harness repair

3a85baa

Wait for observed field geometry within a bounded native trial deadline

023ec2f

Record combined review and fresh native restart and dispatch qualific…

678443e

…ation

chatgpt-codex-connector Bot reviewed Oct 10, 2026

View reviewed changes

Comment thread tools/browser-inspection/redaction.mjs Outdated

Comment thread tools/browser-inspection/frontend.mjs Outdated

Tom-R-Main added 8 commits October 10, 2026 13:36

Predict browser artifact redaction and cross-platform CI remediations

e107417

Redact short URL credentials and allowlist browser inventory counts i…

95be91f

…n CI

Validate frozen visual corpus hashes independently of host rendering

33fdf1f

Keep portable redaction active and isolate macOS-only evaluation probes

14fb64f

Record reviewed CI and browser artifact remediations before fresh Lin…

9306464

…ux qualification

Predict deterministic deadline test boundary synchronization before r…

6dd8164

…epair

Trigger deadline tests at observed input boundaries rather than host-…

e95a852

…speed estimates

Close remediation hypotheses and hand final reviewed source to Linux CI

fd62f6c

Tom-R-Main merged commit 8c9478f into main Oct 10, 2026
1 check passed

Tom-R-Main deleted the codex/h4-thin-slice branch October 10, 2026 18:00

Navigation Menu

Repository navigation

Make Footwork actions verifiable and recover scoped native appends after restart - #1

Make Footwork actions verifiable and recover scoped native appends after restart#1
Tom-R-Main merged 288 commits into
mainfrom
codex/h4-thin-slice

Tom-R-Main commented Oct 10, 2026 •

edited

Loading

Uh oh!

chatgpt-codex-connector Bot commented Oct 10, 2026 •

edited

Loading

Uh oh!

chatgpt-codex-connector Bot left a comment

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

1 participant

Conversation

Tom-R-Main commented Oct 10, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

chatgpt-codex-connector Bot commented Oct 10, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Codex Review Summary

Uh oh!

chatgpt-codex-connector Bot left a comment

Choose a reason for hiding this comment

💡 Codex Review

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

1 participant

Tom-R-Main commented Oct 10, 2026 •

edited

Loading

chatgpt-codex-connector Bot commented Oct 10, 2026 •

edited

Loading