Feature plan: Step 1 · Step 2 · Step 3 · Step 4

Effort study: railspike-effort / v3/feedback_work_restart; host ledger. Standing P2 enrollment.

Extended estimates: enabled

Branch: feedback_work_restart, based on remove-work-tracking (81837cf). Planning commit 6e1909f.

Cycle estimate: 0.60

Step 4 entry projection

2026-11-04; Claude, single exposed rater; R1 not re-verified in this session. 0.60, reaffirming Step 3. Stage identities and original stage-work estimates are copied from Step 3; copying is not a new rating.

Stage 1 - Item format and check command

Original estimate: 0.12 (unchanged)

• Changes:

  • Spec feedback_item_record_v1.md: headers ID, Title, Kind, Workflow, Created, optional Evidence, Estimate, Aliases, Duplicate-Of, Superseded-By; body sections Request:, Sources:, Done when:, Notes:. Sources are body lines because the Key: value parser rejects repeated headers (Step 3 risk, resolved as planned).

  • src/ForumRewrite/Feedback/FeedbackItem.php, FeedbackItemCheck.php (reuses GenericTextRecordParser); scripts/feedback.php check [--dir=]; ./v3 feedback dispatch and usage line.

  • Concept entry feedback_items.md.

• Verification:

  • ./v3 test --smoke FeedbackItemCheckTest: 429 run, 429 passed (7 new tests: valid item, malformed and unknown headers, ID vs filename and kind, invalid values and missing sections, missing duplicate target, shared source, command exit code). Full suite not run.

  • ./v3 feedback check on the empty folder: exit 0.

• Notes:

  • Deviation from Step 2 wording: a post shared by two items is listed as Review: and does not fail the check. The October 30 summary splits combined posts into separate outcomes on purpose, so a shared source is often correct. Malformed files, bad values and missing duplicate targets still fail.

  • Duplicate IDs cannot occur among loadable files because the filename must equal the ID; the cross-file duplicate check is kept as a guard.

  • Retrospective stage-work assessment: 0.12, as estimated; the shared-source decision was the only surprise.

Stage 2 - Generated index

Original estimate: 0.06 (unchanged)

• Changes:

  • src/ForumRewrite/Feedback/FeedbackIndex.php: open items by kind (features, bugs, other), ordered in-progress, blocked, planned, ready, untriaged, then ID; closed items in one table with their outcome and duplicate or superseding target; table pipes escaped; atomic write via a temporary file and rename.

  • ./v3 feedback index [--dir=] writes README.md beside the items folder, refuses to write while check reports problems, and reports Unchanged when the output is identical.

  • Generated docs/feedback/README.md (empty until Stage 3); concept entry updated.

• Verification:

  • ./v3 test --smoke FeedbackItemCheckTest: 430 run, 430 passed (new index test: refusal on problems, grouping and order, escaping, closed outcome links, unchanged rerun). Full suite not run.

• Notes:

  • Retrospective stage-work assessment: 0.06, as estimated.

Stage 3 - Pilot migration of two register entries

Original estimate: 0.06 (unchanged)

• Changes:

  • PF-20261021-F01 (reply focus, the entry with the longest history) and PF-20261021-B03 (saved draft starts collapsed).

  • Register sections for both now start with a Current item: link; their fields stay as migration history.

  • Spec: file references in Notes: are repository-relative paths.

  • Spec and concept examples now use the illustrative ID PF-20260101-F01; the earlier example collided with the register's real PF-20261103-F01.

• Verification:

  • Side-by-side review against the register, field by field: alias, workflow state, delivery evidence and date, request, report link, inherited estimate, recommended workflow, done-when, next action, implementation and checklist links, source-code pointer, historical triage link. All carried over; the F01 2026-10-23 investigation stays in the register and is linked by anchor.

  • All linked paths and the checklist anchor exist; ./v3 feedback check: 2 items, 0 problems; index regenerated.

• Notes:

  • Pilot outcome: the format holds; long investigations belong in linked history, not in the item. No change needed before Stage 4 beyond the path convention.

  • Retrospective stage-work assessment: 0.06, as estimated.

Stage 4 - Migrate the remaining register entries

Original estimate: 0.08 (unchanged)

• Changes:

  • 15 item files: PF-20261021-F02 to F11, B01, B02, B04, B05, and PF-20261103-F01. Converted with a scratch script outside the repository, then checked and hand-corrected.

  • Hand corrections: F02 keeps both its inherited 0.03 and the separate source-informed 0.10; F04 keeps its locally delivered scope and approved-cycle link and has no Done when: (the register gave none, so none was invented); PF-20261103-F01 records the 0.80 Step 1 projection as its estimate with provenance and drops the October 21 triage link it never had; F01 regains the related cycle's boundary link.

  • Every item's Notes: starts with the register status line at migration (review date, evidence limits) and ends with a link to its register section.

  • Register: superseded banner pointing at the items and index; every entry starts with a Current item: link. Its investigations and the related-reports note stay in place.

  • production_feedback_management.md now starts from the feedback index and items; ID, workflow, evidence, closure and correction rules kept, with the O kind and the item-file spellings of states added.

• Verification:

  • Script comparison of register and items: 17 IDs, 17 items, none missing; every zenmemes.com report link in each entry appears in its item; every document link appears, except the new Current item: links themselves.

  • ./v3 feedback check: 17 items, 0 problems, 2 to review. Both are recorded splits: F08, F09 and F10 share one reply; F02, B01 and B02 share another.

  • ./v3 feedback index regenerated: 17 open, 0 closed.

• Notes:

  • Retrospective stage-work assessment: 0.10. Higher than 0.08 because the three irregular entries and the lost status lines needed a second pass.

Stage 5 - Seed from the October 30 summary

Original estimate: 0.15 (unchanged)

• Changes:

  • Proposal list (16 attach, 74 create, 3 for the maintainer, 4 skip) shown before writing; the maintainer approved it 2026-11-04. Other 01 created; Other 06 and 07 skipped by the stated default.

  • 16 attaches: each matched exactly one register item by source and outcome. Missing sources added, alias 2026-10-30/<row> added, and the summary's row text and status appended to Notes:.

  • 75 new items PF-20261030-*, numbered by summary row (53 features, 21 bugs, 1 other), workflow untriaged, estimate only where the summary gave one, titles written for the proposal, request wording preserved.

  • docs/feedback/HISTORY.md: what was folded in, the six skipped rows with reasons, and the inventories not yet reconciled (October 18 and 21 summaries, imported feature requests). The index links it when present.

• Verification:

  • Script coverage check: every source in the 91 non-skipped rows appears in an item; every one of those rows has its alias.

  • ./v3 feedback check: 92 items, 0 problems, 12 to review. All 12 are splits the summary made deliberately (for example one long agent report produced ten API items).

  • ./v3 test --smoke FeedbackItemCheckTest: 430 run, 430 passed (index test now also covers the history link). Full suite not run.

• Notes:

  • Deviation: Step 3 said the 10-18 and 10-21 summaries would be listed as unreconciled "in the index"; they are listed in HISTORY.md, which the index links, to keep the generated file free of hand-written text.

  • Retrospective stage-work assessment: 0.15, as estimated; the summary's own merges made matching straightforward.

Stage 6 - Triage procedure in docs/agents

Original estimate: 0.08 (revised in Step 3 from 0.07 for the docs/agents/ location; unchanged since)

• Changes:

  • docs/agents/README.md: index of agent procedures, stating that runtime prompts stay in prompts/.

  • docs/agents/feedback_triage.md: window, discovery, matching, create/attach/skip proposals, write, check, index, recovery. Carries over the checklist's attribution, context and qualifier rules.

  • zenmemes_feedback_automation_checklist.md marked superseded; minimal_feedback_triage.md and the concept entry point at the items and procedure; AGENTS.md gains an "Agent procedures" section pointing at docs/agents/.

• Verification:

  • Every relative link in the touched documents resolves (the two placeholders in the minimal triage template excepted); git diff --check clean.

  • Rerun behavior on a covered window: the Stage 5 coverage check shows every catalogued October 30 source already listed in an item, so step 3's "already a source" rule proposes nothing for them. A live dry run needs the maintainer's import and is part of Stage 7.

• Notes:

  • Not verified here: a full agent run on real imported posts. This workspace has only the three-post test fixture, not the zenmemes.com import.

  • Retrospective stage-work assessment: 0.08, as estimated.

Stage 7 - Acceptance on the maintainer's WSL checkout

Original estimate: 0.05 (unchanged)

• Evidence so far (2026-11-04, maintainer's WSL checkout, the maintainer's agent):

  • After a fresh ./v3 import-instance, the agent found the procedure, read 377 posts from the local import (window from 2026-10-30 local midnight), wrote nothing, and returned proposals: posts 17 attach, 147 create, 213 skip; outcomes 118 new items (77 features, 32 bugs, 9 other) and 5 existing items updated.

  • The two 2026-11-03 "summary field" threads became one proposed item (PF-20261104-F40) with both sources.

  • Already-sourced October 30 posts were skipped as covered, so the overlap with the seeded window produced no duplicates.

• Applied by the maintainer in 1dd5f66 (2026-11-04): 118 new items, 5 attaches, 1 status update; proposals kept as docs/feedback/triage/2026-11-04_proposals.txt with Status: applied. The maintainer applied the proposals as written, without the suggested merges and skips.

• Catch-up step exercised on real history: found no trailers, reviewed PF IDs in 445 changed plans, and proposed one status update (PF-20261103-F01: merge recorded, deployment still pending), which was applied.

• Verification on the pushed branch: ./v3 feedback check 210 items, 0 problems, 21 to review (shared sources from split posts); ./v3 feedback index reports unchanged. Script check: all 164 posts classified create or attach are listed in an item, so a rerun of the window skips them. The rerun itself was not performed as a separate agent run.

• Notes:

  • Enhancement request (maintainer, during this stage): proposals must be kept in the repository, not /tmp. Implemented in the procedure: proposals go to docs/feedback/triage/<intake date>_proposals.txt, are edited in place during review, marked applied or rejected, and committed with the item changes; HISTORY.md and a triage/README.md point to them. Links checked.

  • The run's chat summary mixed post and outcome counts; the procedure now asks for both, labeled.

• Scope change (approved by the maintainer 2026-11-04, during this stage): completion tracking. Not in the Step 2 scope, which said only that status changes are file edits.

  • AGENTS.md: search items before a fix, name the ID in the plan, add a Feedback: PF-... trailer per addressed item, update the item in the same commit (done only with completion checks met).

  • Triage procedure gains step 2, "Catch up on completed work": scan git log since the previous applied run for trailers and PF IDs, propose status updates for items left unchanged, never infer completion from unnamed commits. Proposal headers count status updates.

  • Management guide and concept entry describe the trailer and the catch-up.

  • Verification: git diff --check clean; the procedure's section numbering and cross-references read in order.

  • Correction after the run: the triage commit carried Feedback: trailers for the items it edited. Trailers now mark work on an item only; AGENTS.md and the procedure say triage commits carry none, and the catch-up scan ignores commits that only add a proposals file.

• Retrospective stage-work assessment: 0.10, against 0.05. Higher because of the two in-stage additions (proposals in the repository, completion tracking).

Cycle outcome

• Completion contract met on the maintainer's checkout: import, agent triage with review before writing, apply, check, index, and rerun coverage. The work-tracking removal (81837cf) is included in this branch.

• Release condition not yet met: the branch is not merged to main.

• Retrospective cycle assessment: 0.65 against the 0.60 projection; the overrun is the two scope additions in Stage 7.

• Follow-ups: move existing development prompts into docs/agents/; reconcile the October 18 and 21 summaries and the imported feature requests (docs/feedback/HISTORY.md); the parked work page, if wanted, as a view over items.