The fifteen-issue backlog sweep (0.21.0) - #96
Merged
Conversation
…ague /scope + /think + /challenge) Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01QtUixLUjgEt51PXS9TPVt3
…evague /spec-to-plan) Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01QtUixLUjgEt51PXS9TPVt3
…79, #91, t16) Steps 1-2 of the /scope method now instruct subagent fan-out for a broad survey (5+ candidate surfaces), one read-only exploration subagent per surface or tight cluster, defaulting to a smaller tier (sonnet). 4 or fewer surfaces still explore inline, and the no-wizard escape for small ideas is unchanged. The load-bearing rule survives explicitly in both step 2/4 and the Hard rules section: subagents explore and report only, never running a devague move — the main agent alone runs capture/scope/question/park so provenance and the anti-fabrication contract stay in one place. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01QtUixLUjgEt51PXS9TPVt3
`_require_target` only checked the plan's stored coverage-target snapshot, frozen at `plan new`/last `plan converge` time. But `plan status`/`converge` re-derive targets from the LIVE source frame, so a frame that legitimately grows a confirmed claim mid-run left `status` recommending a cover that `cover`/`task --covers` immediately refused as "unknown coverage target" -- the exact #90 repro. `_require_target` now falls back to the live frame when a target is absent from the stored snapshot, and persists the refreshed snapshot on success so the stored copy catches up without a separate `plan converge`. Decision (park v4 / plan risk r2): when the source frame has itself regressed below its own convergence gate, `_live`'s regression error is let through as-is rather than reworded into "unknown coverage target" -- an unverifiable target should say why, not blame the wrong thing. A target already known to the stored snapshot never touches the live frame at all, so it keeps working through a frame regression exactly as before.
… (t2) store.load and plan_store.load parsed via from_dict before checking schema_version, so a genuinely newer-schema file crashed with a raw TypeError (HardQuestion(**q) / Vagueness(**v) reject unexpected kwargs) instead of the intended fail-closed IncompatibleSchemaError. Both loaders now read the raw JSON, check schema_version first, and only then build the domain object. HardQuestion/Vagueness construction in frame.from_dict is now tolerant of unknown keys the same way Claim already is, so a future field lands safely under a version bump instead of crashing same-version loads that happen to carry it. Bumps frame.SCHEMA_VERSION 3->4 (reserved for t4's HardQuestion resolution field) and plan.PLAN_SCHEMA_VERSION 3->4 (reserved for t9's per-target deferral state) - this task only hardens load order/tolerance, it does not add either field. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01QtUixLUjgEt51PXS9TPVt3
…, t1) Add md_safe_text() to render/_md_safety.py: wraps underscore/dunder identifiers (_read_file, __init__.py) in code spans per the #87 issue comment's preferred fix (fixes MD037 and MD050 in one move, reads better than backslash-escaping), and backslash-escapes the remaining control characters (*, [, ], a stray backtick, a leading #) as the fallback. Text already inside a code span is left byte-for-byte untouched and double application is a no-op (c32/h25). No renderer call sites yet — those land in t3/t9/t13.
Add `split-plan --write`, producing docs/plans/<created-date>-<slug>-split.md alongside the exported plan-md: real per-task summaries, instructions, and acceptance criteria from `devague plan waves --json`; a Task assignments (Owner/Model) table the script reads back on re-run so a human's edited assignment survives regeneration; and the End state section. Re-running overwrites the same dated path in place. Artifact-only per decision c25 — no plan schema change, no new devague CLI verb; devague plan waves/show/ deliverables stay read-only. SKILL.md documents the flow.
…48, #52, t4) interrogate <cN> --resolve <qN> [--decision "<text>"] adds Frame.resolve_hard_question and a resolve mode on the interrogate CLI parser (mirrors park --resolve), fixing the permanent convergence deadlock where nothing ever set HardQuestion.resolved. The blocking-question gate also skips rejected claims, and suggest_move now names the shipped move instead of the old dead-end "capture/confirm the resulting claim" hint. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01QtUixLUjgEt51PXS9TPVt3
…ks (#88, t13) devague summary rendered every task on the plan -- including rejected ones -- into both Planned Work and the Actual Delivery table, so a plan carrying 19 confirmed and 68 rejected tasks (rebuilt twice after scope changes) emitted 87 undifferentiated rows. A rejected task paired with `<fill: status>` invites recording a planning decision as a delivery failure, which is a different fact. Planned Work and Actual Delivery (and the --json summary_data equivalent) now iterate confirmed tasks only, with a single line preserving the rejected count without padding either list ("N tasks were rejected during planning -- see `devague plan show`"). A proposed task -- neither the confirmed contract nor an explicit rejection -- is excluded from both lists and from the rejected count, so an open decision is never folded into a closed one. The --pr wave map is unchanged (dependency_waves already excludes rejected tasks), pinned here with a dedicated regression test. Also applies the #87 md_safe_text escaper at every verbatim interpolation site in this renderer (task summaries, deviation what/reason, announcement/after-state text). Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01QtUixLUjgEt51PXS9TPVt3
…, t9) Add `devague plan defer <target-id> --reason "<text>"` (+ `--undo` to reverse it) so a plan intentionally scoped to early milestones is not forced to fake coverage of a later-milestone target just to satisfy the gate — the exact perverse incentive #85 reported (a reviewer catching a task that "claims coverage ... but only installs and reports the counter"). - plan.py: CoverageTarget gains deferred/deferred_reason (schema v4, reserved by t2); Plan.defer_target/undefer_target mirror resolve_risk's fail-closed contract (unknown id / already-deferred / not-deferred all raise). - plan_convergence.py: _missing_coverage excludes deferred targets from blockers; _parked_items now also lists each deferred target labeled distinctly ("deferred: coverage target cN ... — <reason>") so `plan status`/`converge --json` can tell "deliberately deferred" apart from "not yet covered". - cli/_commands/plan.py: new `defer` subcommand + PLAN_MOVES entry; _live/_live_frame_and_targets now carry persisted deferral state across every live-frame re-derive (targets_from_frame builds fresh CoverageTarget instances with no memory of prior deferrals — without this merge, converge/export/status would silently drop a recorded deferral on the next call). - cli/_status.py: emit_status now prints parked_items in text mode too (a pre-existing gap — JSON already carried them) so `plan status` visibly shows deferred targets, not just JSON. - render/plan_md.py: new "Deferred targets" section naming each with its reason; applies the t1 escaper (md_safe_text) at every verbatim site (task heading/instruction/acceptance, risk text, announcement, title) — closes the MD050 regression from the #87 comment for underscore-bearing task text. Verified against real markdownlint-cli2 (0 errors) and an end-to-end `devague` CLI smoke run in a scratch dir (defer -> converge -> export -> status -> undo), plus the full pytest suite (760 passed) and black/isort/flake8. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01QtUixLUjgEt51PXS9TPVt3
… verbatim text (#93, #49, #83, #87, t3) spec_md.py's park-kind filter only ever surfaced follow_up/out_of_scope, silently dropping every open unknown_nonblocking/unknown_blocking park — exactly the residual-risk kind that legitimately coexists with a converged frame (#93, #49). It now renders every open park kind, grouped and labeled. Hard questions previously carried no resolved marker and rendered regardless of parent-claim status, so a resolved question read as an open blocker and a rejected claim's question still leaked into the artifact (#49, #83). Both are fixed: resolved questions get a "(resolved)" marker, and any claim.status == "rejected" drops its hard questions entirely. A scope entry whose seeds cite a claim that was later rejected now renders a visible "(rejected)" marker instead of a bare dead id (#84's fourth acceptance criterion, c33/h26). Every verbatim interpolation site now composes the t1 escaper (md_safe_text) with the existing autolink_urls/heading_safe passes via new _safe()/_safe_heading() helpers, fixing MD037/MD050 on underscore/dunder identifiers throughout the exported spec (#87). Flips the pinned test_render.py:301-306 (nonblocking parks now render) and adds regression coverage for all four acceptance criteria, including a real markdownlint-cli2 double-export byte-stability check. Re-exporting docs/specs/2026-07-28-issue-backlog-sweep.md (a live corpus mixing backticked and bare identifiers) produces the expected one-time presentational diff — committed alongside, per park v5. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01QtUixLUjgEt51PXS9TPVt3
… escaper (#87) Proves the t1/t9 escaping fix end-to-end: three pre-existing MD037 errors in a committed artifact from PR #81 clear on re-export, with no source text change. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01QtUixLUjgEt51PXS9TPVt3
… t5) Frame.reject cascades a claim's rejection onto its still-live honesty conditions (status -> rejected) and reports both cascaded honesty and hard-question ids so the CLI can echo what it took with it (`c21 -> rejected (also rejected: h3, q1)`). The cascade fires only on the transition into rejected, so re-rejecting an already-rejected claim reports nothing (idempotent, no double-reporting). _assumption_warnings now skips rejected assumptions (only a still-proposed one is actionable "unconfirmed"), and devague review naturally stops listing an honesty condition once its parent claim's rejection cascades over it -- no renderer changes needed, since proposed_honesty() already filters on status == "proposed".
`plan task --dep` and `depend <tN> --on <tM>` now refuse a self-cycle (the about-to-be-assigned/naming-itself task id) and an unknown task id at authoring time, with an actionable hint, instead of silently recording a broken edge that only surfaces much later as a bare `dependency cycle: tN -> tN` at `plan converge`/`plan waves`. `depend --remove` is untouched by the new checks (it must still repair a dangling dep predating this fix), and the plan_convergence gate for multi-task cycles is left alone — this is creation-time feedback, not a replacement for it. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01QtUixLUjgEt51PXS9TPVt3
#92, t14) Adds a read-only derivation (devague/contested.py) joining a frame's confirmed claims to any approved `devague deviate` record naming them in --affects, since the deviation ledger knew a claim had been contradicted by execution but nothing rendered that back-reference (the spec, `show`, and `status` all stayed silent). Per the #92 maintainer ruling the spec itself is never rewritten -- this only ever derives a pointer at render time: `export` gets a rich per-claim marker (announcement blockquote, Requirements, and every generic claim section), `show`/`status` gain a summary "contested: ..." line in both text and --json, and the plan engine's own status is untouched (no "contested" key at all). The join crosses the frame/plan boundary the instruction called out: frames carry no reverse pointer to the plans seeded from them, so every plan slug is enumerated and filtered on frame_slug before its delivery ledger is read. Fails open throughout -- a plan or delivery file that is missing, truncated, or declares a newer schema degrades to "no markers from that source" plus a stderr diagnostic, never a crash or a refused export; zero frame/plan/ delivery mutation. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01QtUixLUjgEt51PXS9TPVt3
plan confirm/reject now accept N ids (nargs="+") and apply them transactionally — every id is validated against the plan first, and if any is unknown nothing is changed, matching the frame-side contract (confirm.py's _run). Argument errors raised inside a devague plan <move> parser now hint at `devague plan explain <move>` instead of the generic `<prog> --help`, scoped so top-level and other flat-verb errors keep their existing hint unchanged.
New `devague amend <cN> --text/--kind [--reason]` and `devague scope --amend <sN> --finding` edit a claim or scope entry in place — same id, same honesty conditions/hard questions/instruction, same inbound scope-entry seeds. The only prior route (reject + recapture) churned the id and left scope-entry seeds dangling. Amending a confirmed claim flips it back to proposed and echoes the transition, mirroring the `interrogate --instruction` precedent; origin is never touched (no flag reaches it). A lightweight `Claim.revisions` trail (text/kind/reason) records what an amend superseded, added without a schema_version bump since `from_dict` loads it tolerantly.
Frame.add_scope_entry validates a seed against find_claim first and the new find_hard_question second, so a claim-attached hard question id (q*) records alongside claim ids (c*) -- unknown ids of either kind still refuse with the existing "unknown seed claim id" error and "run 'devague show'" hint. render/spec_md._seed_label renders a question seed as `(question)`, or `(question, resolved)` once answered, distinct from the existing claim/rejected-claim markers. This closes the "smaller, related gap" in issue #84: the /scope skill's own routing table sends a "genuinely unknown, needs a user decision" finding to the `question` move rather than `capture`, so a scope entry recording that finding previously had no way to link its provenance.
Add `Plan.amend_risk` and `devague plan risk --amend RID --text "<corrected>"` so a risk whose prose names a task id that rotated (rejected + recreated during a scope change) can be corrected in place instead of resolving it just to record a duplicate. Preserves id, kind, task link, and resolution state verbatim -- a resolved risk stays resolved after its text is fixed.
interrogate --resolve, amend, scope --amend (+ --seeds accepting q* hard-question ids), plan defer, plan risk --amend, transactional multi-id plan confirm/reject, live plan cover/--covers validation, the flat reject cascade, and the scope subagent fan-out threshold were all shipped ahead of learn/explain catching up — devague explain amend even 404'd with "unknown move: amend" before this. Also fixes plan.py's cmd_plan_learn, which still named "six operator skills" and omitted challenge from the list. Closes the #52 acceptance criterion: learn/explain document the resolve path. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01QtUixLUjgEt51PXS9TPVt3
…ue sweep (t18) Bump 0.20.1 -> 0.21.0 (minor: feature release) and describe the shipped surface across every teaching doc. - CHANGELOG.md: one 0.21.0 entry mapping each of the fifteen issues to what changed (#48 #49 #52 #79 #82 #83 #84 #85 #86 #87 #88 #90 #91 #92 #93). - CLAUDE.md: new Status entry; working-backwards and spec->plan method lists gain amend / interrogate --resolve / scope --seeds q* / plan defer / risk --amend / multi-id plan confirm / live-frame cover; gate 2 now names its durable split artifact; module inventory picks up amend.py, contested.py, _md_safety.py and plan defer. - README.md: verb lists, the "nothing is deleted to make a gate go green" close-out rule, and the reject cascade. - docs/skills.md: scope subagent fan-out and the split-plan --write artifact; fixes the pre-existing drift that said briefs quote plan show --json / the plan-md (SKILL.md is right: plan waves --json is the single source). - docs/skill-sources.md: version stamps for scope and assign-to-workforce. - docs/spec-contract.md: schema_version 4 on both stores (check-before-parse), ClaimRevision, HardQuestion.resolution, CoverageTarget.deferred, the new moves and their validation errors, plus render-time contracts (md-safety, contested-by-deviation). - docs/llm-guidance.md + think/scope SKILL.md: the stale --seeds claim-ids-only wording, and the three new close-out/correct-in-place rules. Docs and version only — no behavior change.
Adversarial verification of all fifteen issues in the sweep, each checked against its own stated acceptance criteria and — where the issue carried repro steps — by running the reporter's sequence against the built CLI in a scratch dir rather than by reading the tests. Closure map: docs/deliveries/2026-07-28-issue-backlog-sweep-closure-map.md (967 tests green, 98.37% coverage, 0 markdownlint errors including the force-linted .claude/skills/** files). Three defects the per-task agents did not catch, fixed here with tests: - #94 (regression this release introduced): md_safe_text and autolink_urls were composed in OPPOSITE orders by spec_md vs plan_md/summary_md, and both corrupted a URL containing an underscore — the plan order backticked inside the link, the spec order truncated it at the first underscore, silently pointing a committed artifact's link at the wrong address. md_safe_text now carves out URLs exactly as it already carved out code spans, so both orders produce identical intact links. - #49 was only half closed: a resolved hard question rendered "(resolved)" but dropped the recorded decision text, while the parks renderer already rendered its own. The issue asked for "a pointer to the claim/decision that answered them", so the export now renders "(resolved: <decision>)" when interrogate --resolve --decision recorded one. - spec-to-plan/SKILL.md was never swept and still taught `plan reject` as single-id with a shell loop — the exact workaround #86 removed — in the skill guildmaster re-broadcasts to the mesh. Moves table corrected and pinned against `devague plan --help` by a new test. The markdownlint integration test now exports the repo's own real issue-backlog-sweep frame and plan as lint corpus, per the task instruction. Two CHANGELOG factual errors corrected: "Three of those" listed only two workarounds (#87's lint ignores were missing), and the swept-surfaces list omitted spec-to-plan.
…resh clone The contested-marker derivation (#92) reads .devague/deliveries/<plan-slug>.json; without it committed, a fresh clone cannot reproduce the markers. Flagged by t19. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01QtUixLUjgEt51PXS9TPVt3
Eight-section accountability artifact built from the devague summary skeleton: 19/19 tasks delivered, 2 approved deviations (d1/d2, execution-order only), 3 defects found and fixed by t19's cross-task verification, and the honest remaining-work list including the two follow-up issues opened during the run. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01QtUixLUjgEt51PXS9TPVt3
Contributor
Author
|
/agentic_review |
Qodo is busy workingCheck back in a few minutes. Qodo's code review agents are on it. |
- contested.py: extract _delivery_for_frame and _markers_from_delivery so find_contested_markers drops from cognitive complexity 24 to under the 15 allowed (python:S3776). - _md_safety.py: name the file-extension alternation instead of relying on implicit string concatenation in the regex (python:S5799). Behavior unchanged; 967 tests still pass. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01QtUixLUjgEt51PXS9TPVt3
|
PR Summary by QodoIssue backlog sweep (0.21.0): 15 issues — new moves, cascades, and export fidelity fixes
AI Description
Diagram
High-Level Assessment
Files changed (72)
|
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Closes the entire open issue backlog — fifteen issues — in one release, plus one opened and fixed during the run.
Closes #48, #49, #52, #79, #82, #83, #84, #85, #86, #87, #88, #90, #91, #92, #93, #94.
Executed as a single
/assign-to-workforcefan-out of theissue-backlog-sweepplan: 19 tasks, 6 dependency waves, one agent per task per wave in an isolated git worktree, every merge TDD-gated (suite green before and after). No merge was reverted.Why one release
Three of the fifteen were being worked around by hand downstream, and the workarounds were getting expensive:
.devague/frames/<slug>.jsonto setresolved: true, because a blocking hard question could never be cleared through the CLI — nothing in the codebase ever setHardQuestion.resolved, so one--blockingquestion wedgedconvergepermanently (Add a CLI verb to resolve blocking hard questions (converge is permanently blocked without one) #48, Add a CLI verb to resolve a claim's blocking hard question (convergence deadlock; unblocks /think exports) #52).scripts/render_plan.py, with a SHA-256 drift check) because a milestone-scoped plan could not converge — the coverage gate demanded every target, and rewarded tasks that merely named one. Its docstring says to delete it when Milestone-scoped plans cannot converge — the coverage gate rewards tasks that mention a target without delivering it #85 lands (Milestone-scoped plans cannot converge — the coverage gate rewards tasks that mention a target without delivering it #85).docs/specs/**from markdownlint entirely, because exported claim text naming Python identifiers (_read_file,__init__.py) failed MD037/MD050 in their CI (export renders claim text unescaped — underscore identifiers make the spec fail markdownlint (MD037) #87).All three are verified deletable — see the closure map.
What ships
New moves
devague interrogate <cN> --resolve <qN> [--decision]devague amend <cN> [--text] [--kind] [--reason]devague scope --amend <sN> --findingdevague plan defer <target> --reason/--undodevague plan risk --amend <rN> --textassign-to-workforce split-plan --writeSharpened —
rejectcascades onto honesty conditions and unresolved hard questions andconvergestops warning about rejected assumptions (#83);plan confirm/plan rejectare multi-id transactional and plan-group errors point atplan explain(#86);--dep/depend --onrefuse self-deps and unknown ids at creation (#86);cover/--coversvalidate against live-frame targets (#90);summaryscopes to confirmed tasks (#88);scope --seedsacceptsq*ids (#84).Export fidelity — all four park kinds render under
## Open parks, resolved hard questions carry(resolved: <decision>), hard questions on rejected claims are excluded, and a scope seed citing a rejected claim renders(rejected)(#93, #49). Newmd_safe_text()escapes every verbatim render site — presentational only, stored JSON untouched (#87). New read-onlydevague/contested.pymarks a confirmed claim contested by an approved deviation across the plan-slug boundary; the spec is not rewritten, it points forward to the ledger (#92).Both schema versions are now 4, and both stores check the declared version against the raw dict before parsing — previously a newer-schema file died with a bare
TypeErrorfromHardQuestion(**q)instead of the fail-closed error.Evidence
markdownlint-cli2 "**/*.md"— 39 files, 0 errors;.claude/skills/**force-linted — 19 files, 0 errorsdocs/deliveries/2026-07-28-issue-backlog-sweep-closure-map.mddocs/deliveries/2026-07-28-issue-backlog-sweep.mdWhat the verification pass caught
The final task was adversarial cross-task verification, and it earned its place — it found three defects the per-task agents missed:
spec_md.pycomposedautolink_urls(md_safe_text(t))whileplan_md.py/summary_md.pycomposed the opposite order, each documenting its own as correct. Both corrupted an underscore-bearing URL — the spec order truncated it at the first underscore, silently pointing a committed artifact's link at the wrong address. This was filed speculatively as autolink_urls and md_safe_text compose without knowing about each other — a URL containing an underscore can be corrupted #94; it was real. Fixed by carving URLs out ofmd_safe_textthe way code spans already were.(resolved)but dropped the decision text, while the issue asked for a pointer to what answered it.spec-to-plan/SKILL.mdwas never swept — it still taughtplan rejectas single-id with a shell loop, the exact workaround plan papercuts: self-dependency accepted at task creation; plan reject is single-id while frame reject is multi-id #86 removes, in a skill guildmaster re-broadcasts to the mesh.Deviations
Two approved mid-run deviations, both execution-order only, both recorded via
devague deviatebefore resuming:d1— t17 (learn/explain recipes) moved from wave 3 to after wave 5: three of the four surfaces its acceptance criterion names don't exist until waves 4–5, so documenting them in wave 3 meant writing recipes for unshipped verbs.d2— t18 (docs sweep + version bump) moved to last, same root cause.No plan state was mutated — deviate is the marking of the change.
Follow-ups
r4).qNnamespace collision (claim-attached hard questions vs. the durable questions file both mintqN) is documented as a hazard this run; not yet filed.Review map
The delivery summary is the audit trail: every plan task accounted for, both deviations quoted, every delivery claim carrying a resolvable evidence pointer or an explicit
unverifiedmarker.