Spaces:
Sleeping
Sleeping
Commit History
-
docs: three guidance-extraction defects, found by reading the data back
All three surfaced the same way: comparing a stored figure against the quotation stored beside it. That is what the "every guidance figure carries its passage" rule was for, and this is the first time it has been used. Three of 47 margin bands overstate the upper bound by exactly 5.00 points because "50 basis points" was read as five percentage points. The signature is mechanical -- high = low + 5.00 -- and the quote makes it unambiguous, so a validator can refuse a band whose width contradicts the tolerance its own quote states. Revenue extraction is clean. Guidance issued in a 10-K carries the period label Q1FY2025 where the metrics table uses Q12025, so it never joins and drops silently out of any guidance-versus-actual scoring: on NVDA, 8 of 14 rows scored, 2 lost to the label, 4 targeting a Q4 that has no 10-Q at all. Segment rows stop one quarter short of the metrics table, so the newest quarter -- the one a reader opens first -- cannot be decomposed. Also recorded as NOT a d…e4b1b7a -
docs: close both defects in the run record, and correct one classification
Eight of eight acceptance criteria now pass; row status is 88 VERIFIED / 1 CHECK_REQUIRED / 0 LEGACY_UNVERIFIED. GOOGL Q12023 was ours and is fixed: filings.recent holds the last ~1000 submissions of any form, and Alphabet's Form 4 volume had pushed its April 2023 10-Q into the filings.files archive we never read. NVDA Q32024 is reclassified from "defect, follow-up needed" to "the filer does not tag it", which is what the investigation actually shows. Its Q2 FY2024 filing tags no property-or-equipment capex concept at all, so the year-to-date figure the de-accumulation needs did not exist when the Q3 filing was made; it appears only in the next year's comparative, filed nine months later. Using that would subtract a figure restated a year afterwards from an as-filed one. The row keeps its masked free cash flow and its warning, which the spec calls a correct outcome. The wrong label mattered: it would have sent someone hunting through our own code for a gap that belongs to the filer.…58b385d
-
data: re-extract 88 periods, guidance from the calls, segments on six tickers
Row status moved from 41 VERIFIED / 15 CHECK_REQUIRED / 33 LEGACY_UNVERIFIED to 87 / 1 / 1. Structured guidance went from 0 rows to 53, each carrying the passage it came from, with zero parse errors across 88 periods. Segment rows went from 275 on one ticker to 1,288 across all six. Every fallback_non_sec warning is gone. Run record in docs/superpowers/plans/2026-09-01-data-foundation-run.md names every residual gap and labels it. Two are defects with follow-ups: GOOGL Q12023 was never reached because EDGAR caps submissions.filings.recent at ~1000 entries and Alphabet's Form 4 volume pushes that array's floor to 2023-06-27, so the April filing lives in the filings.files archive the ingestion never reads; and NVDA Q32024 cannot de-accumulate its prior year-to-date capex. The 84 statement-line absences are filer facts -- Apple reports no separate goodwill line, AMD tags no current contract liability. GOOGL and TSLA returned 28 rows of NO_GUIDANCE_ISSUED, which is the correct reading of…1f17471 -
docs: implementation plan for the data foundation (sub-project 1)
Seven tasks, each ending in a testable deliverable and a commit. Six are code with TDD cycles; the seventh runs the backfill and writes down what it found, gap by gap. The plan carries the diagnosis rather than restating the goal: the parser window and the inert paragraph filter (task 1), the three-way ambiguity in a null guidance value (task 2), Alphabet's absent GrossProfit concept and NVIDIA's mid-history capex switch (task 3), eight concepts already on the wire (task 4), and the re-extraction mode that neither delta nor --full can provide (task 5). Task 7 requires every residual gap to be labelled either "filer does not tag it" or "defect, follow-up needed", so a shortfall cannot be absorbed into the run record as if it were normal. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_014WiLV6MBBYAD5sWnv5Xqc5399f59d -
docs: correct Workstream C — the two fallback causes, diagnosed
The spec asserted both fallback_non_sec families were missing candidate concepts. Half wrong, and the wrong half mattered. All 8 gross-margin failures are GOOGL, and Alphabet does not tag GrossProfit at all -- the concept is absent from its companyfacts payload. It publishes Revenues and CostOfRevenue and leaves the subtraction to the reader. So this one needs a derivation, revenue - cogs, gated on both facts sharing an accession and a period, recorded as derived_difference exactly as free_cash_flow already is. Not a tuple extension. All 7 free-cash-flow failures are NVDA, and there the original reading holds: capex moved from PaymentsToAcquirePropertyPlantAndEquipment (through FY2020) to PaymentsToAcquireProductiveAssets (FY2021+), and only the first is tried. Every failing row is FY2024 or later. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_014WiLV6MBBYAD5sWnv5Xqc56cb4a3a -
docs: analysis-depth roadmap and the data-foundation spec behind it
The dashboard review turned into a data review. Measured, not assumed: structured guidance is 0/89 rows, 33 of 89 metric rows carry no SEC provenance and are masked whole-ticker, statement lines sit at 56/89, and segments exist for one ticker of six. Root cause found for the guidance blank and validated by re-running the parser: guidance_parser truncates the transcript to 4,000 characters, and its paragraph filter is inert on transcripts (34 newlines in 50,000 chars yields one paragraph), so it reads the operator's introduction. NVDA's outlook section begins 33.9% into the call. Widened, the same stored transcript yields Q42026 revenue 63.7-66.3B and gross margin 74.4-75.4%. Also verified live: companyfacts already carries every re-declaration of a period, so restatement detection is available with no new data source -- 42 divergent period-concept pairs on NVDA and TSLA, 0 on AAPL revenue -- provided a materiality threshold separates real revisions from rounding changes. The roadmap…090e385 -
docs: design spec for falsification conditions (sub-project D)
The storage layer already exists and has no producer: forecast_store is built and tested, data/forecast_log.jsonl does not exist, and nothing in the repository calls append_forecast. What is missing is a thing that registers conditions and a thing that resolves them. The design turns on a closed metric vocabulary in three tiers -- filed columns, derived signals, segment lines -- so resolution is arithmetic over the store rather than a judgement. The model proposes; validation refuses anything it cannot evaluate for that ticker today. A probe against the live store confirmed all three tiers resolve and that an unknown segment member fails closed. No probabilities: those need calibration to mean anything, and calibration is the sub-project this one exists to feed. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>e5e550f -
chore: rename the project to Readthrough
Amplegest is an existing Paris asset manager, founded 2007, EUR 5.2bn under management, with a European equity fund range — the same buyers this tool is built for. Shipping under that name carried a trademark exposure and read, to the intended audience, as either a competitor's product or a mistake. Readthrough is the buy-side's own term for the implication a disclosed fact carries for the numbers, which is exactly what this produces: it reads the filing, reports what changed, and quantifies the effect only where the company quantified it. The codebase already spoke the word — the brief's lead panel is keyed read_through_label. Renames the wordmark, page title, intro copy in four languages, the plotly template, the SEC and prospection user agents, and the docs. Nothing behind the name changes. Also ignores .tmp_* scratch runtimes, which have no business in a commit. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>4d22778
-
Merge branch 'sdd/prospection': buy-side prospection sub-system
Turns the AMF open-data registry of authorised asset managers into sourced, GDPR-compliant email drafts, each carrying a real research brief on a position the firm actually holds. Runs end to end with every paid provider disabled. 14 tasks, each reviewed and fixed; a whole-branch review then closed three blocking defects the task-scoped reviews could not see. Execution decisions are recorded in docs/superpowers/plans/2026-08-29-prospection-buyside-decisions.md. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_016XzekHxqiXutBW4gw3Arsxdf05467 -
docs: preserve the prospection execution decision log
86 rulings made during subagent-driven execution, with the reasoning and the cost if each is wrong, plus the pre-flight conflict scan and the deferred findings. A ruling made on someone's behalf that dies with its scratch directory was made in secret. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_016XzekHxqiXutBW4gw3Arsx5a6c4fc
-
docs: fix the real bootstrap block - pass 1 gated on a value it cannot have
Moving holdings discover to Tier 2 addressed the second link and left the first. Pass 1 promoted only accounts with fit >= 25, but fit reaches 25 only with an EDGAR overlap, read from confirmed holdings that do not exist on a fresh database. So pass 1 promoted nobody, contacts enrich filtered on tier 1 and saw none, and the named-contact gate then closed Tier 1 for good. Pass 1 now admits any account with declared equity exposure - which is what choosing who to pay to enrich actually means - while MIN_FIT still gates final Tier 1. The spec was wrong here too, not only the plan. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_016XzekHxqiXutBW4gw3Arsx82ea202 -
docs: rule on the Task 14 findings - the Tier 1 bootstrap circularity
Tier 1 needs a confirmed holding; holdings discover looked only at Tier 1; on a fresh database that promotes nobody, ever. The integration test hid it by hardcoding the exact value the real pipeline cannot produce. Discovery now starts at Tier 2, which only needs declared equity exposure, and the README explains why score --final runs twice. The test now replays the CLI's own _score_inputs and asserts Tier 1 opens only after a holding is confirmed. Also corrects CLAUDE.md: demo build does call the app's shared LangGraph agent. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_016XzekHxqiXutBW4gw3Arsx43ff59b -
docs: separate enforced refusals from human process rules in the prospection README
095f50e -
docs: rule on the Task 13 CLI findings
A contact with no address - exactly what the LinkedIn provider produces, and not generic so nothing upstream skipped it - produced a written .eml headed "To: None". render_eml now refuses an empty recipient and export degrades per touch instead of stranding every remaining one. An unconfigured sender identity crashed drafting mid-loop; it is now checked once, before the loop. Both --tier flags are wired through rather than silently ignored. And the suppression gate finally has a test: deleting it entirely used to fail nothing. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_016XzekHxqiXutBW4gw3Arsxda54928 -
docs: rule on the Task 12 demo-builder findings
A pdf_builder failure aborted the whole ticker-fallback loop instead of trying the next confirmed holding, and a crash between the file write and the audit row left a PDF on disk with nothing describing it. Both now degrade and record. Also narrows ACCEPTABLE_STATUSES to COMPLETE, the only value the brief producer emits (agent/schemas.py:898) - the test fixture had been asserting on VERIFIED, a value the product cannot produce, so the suite was not testing the contract. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_016XzekHxqiXutBW4gw3Arsx7ab8ab9 -
docs: the RGPD placeholder must appear exactly once (T11)
56e5bf4 -
docs: rule on the Task 11 validator findings
Two Critical, both reproduced live. _normalise stripped trailing zeros from integers, so "100" became "1" and a draft claiming a 5.4% position was 100% of the portfolio passed with zero violations; and the offer's price string was tokenised into the allowed set, licensing any single-digit figure. The offer text is now removed from the body before the scan rather than added to the set, and magnitude survives normalisation. Separately, anything the model wrote after the RGPD placeholder escaped the number check entirely. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_016XzekHxqiXutBW4gw3Arsxf44ff44 -
docs: drop a duplicated pytest import in the Task 10 test
b737466 -
docs: rule on the Task 10 compliance findings
4c77ac8 -
docs: word-bound the holdings reject list so TotalEnergies survives (T9)
9bca060 -
docs: rule on the Task 9 holdings findings
1df4f1e -
docs: sync Task 9 with the retrieved_at column and the PDF fixture gitignore rule
04387c6 -
docs: close the unclosed-script leak and restore feminine role forms (T8)
6b5d570 -
docs: rule on the Task 8 signal findings
"data" is a substring of "mandataire", which sits in the legal footer of every French asset manager, and "recherche" is the label of any search box - so the careers scan would have minted a 15-point hiring signal on nearly every site, and since signals are 25 of 90 points in pass 1 that does not add noise, it reorders Tier 1. Matching moves to word boundaries over text with script and style bodies removed, and now requires both a hiring word and a role word. The recent-licence window is derived from the signal's own TTL instead of being a second, slightly different copy of "eighteen months". Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_016XzekHxqiXutBW4gw3Arsx03f8add -
docs: rule on the Task 7 provider findings
The Explorium handoff had no link between a response file and the request it answered, and stamped retrieved_at from the read clock - so a six-month-old cached answer was reported as freshly retrieved, on the very field whose job is to say otherwise. The request now carries a fingerprint the response must echo, and the collection date comes from the response or the file mtime, never the clock. Dropcontact stops taking emails[0] blindly when a qualified address sits later in the list. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_016XzekHxqiXutBW4gw3Arsxda89924 -
docs: the LinkedIn test must name a vendor to reach the path it checks
bab59ea -
docs: rule on the Task 6 waterfall findings
The generic-mailbox fallback was unreachable: run_waterfall only ever calls find(account), so a firm with no named director would have got no contact row at all. It is now unconditional, which also repairs the ABC signature mismatch. Dedup keys are indexed so two anonymous records from one provider cannot collide and silently lose one. The Produces line drops a constructor parameter the reference code never had, rather than growing a dead one to match it. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_016XzekHxqiXutBW4gw3Arsx5473d39 -
docs: note where the usual-name convention is applied (Task 6)
948901d -
docs: rule on the Task 5 scoring findings
Tier 2 now honours the spec's exclusion rule (fit produit = 0 is excluded however high the total) - a firm with no listed-equity book cannot be a lead for a product that reads listed-equity filings. score_reach stops reporting "no contact" about accounts that have one, score_access records a reason on every branch, and the tier cap gets a secondary sort so the account sitting at the 40 boundary does not move between runs. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_016XzekHxqiXutBW4gw3Arsx7bb36d4 -
docs: rule on the Task 4 findings
INSEE NN means headcount unknown, not zero employees - dropped from the map so it yields None. Compound given names are kept whole: reducing "Marie Christine" to "Marie" invents a different real name, and the usual-name convention belongs in the pattern guesser, not in a stored fact. HTTP failures are counted and reported by universe refresh rather than swallowed in the leaf function. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_016XzekHxqiXutBW4gw3Arsx3ee72c9 -
docs: rule on the Task 3 registry findings
A 304 with no cached file would have written a 0-byte CSV and reported an empty prospect universe as a success — silent-empty is what the fail-closed policy exists to prevent, so the loader now raises. Also adds the fixture row that makes the third exclusion audit code reachable. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_016XzekHxqiXutBW4gw3Arsx498321b -
Merge branch 'sdd/attention-brief' into attention-dashboard
d705e12 -
docs: rule on the Task 1 provenance gap
holdings gains retrieved_at: as_of is the position date printed in the fund's report, not the date we read it, so nothing distinguished a holding scraped this morning from one scraped months ago. signals gain nothing — observed_at and url already are the retrieval timestamp and the source. Spec prose amended to name the actual per-table fields instead of contradicting its own model. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_016XzekHxqiXutBW4gw3Arsx1e5f13b -
docs: pre-flight rulings on the prospection plan
Four conflicts found scanning the plan before execution, two load-bearing: accounts never stored type_activite, so fit scored 0 for every account and no account could reach Tier 1; and directors were fetched then discarded, so the free pattern provider yielded nothing. Both now persist, with tests that lock them. Also: latest_demo tie-break on id, and a draft fixture whose clock times tripped the unsourced-number check. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_016XzekHxqiXutBW4gw3Arsxe086d2d -
fix: close four review minors on the attention-ordered brief
- tone_strip._reading split on the Unicode arrow only while _SERIES_RE already accepted "->" as well, so an upstream ASCII normalisation would have kept the rate series working while the plain-language reading silently vanished. Both halves of the parser are now tolerant. - test_kpi_footnote_html_is_byte_identical_without_source_url copied the footnote composition out of _render_quarter_kpis and asserted on the copy, so changing the separator left it green while it claimed to be a regression net. Replaced with a spy on verdict.st.markdown driving the real renderer -- confirmed to fail when the separator is changed, which the old test did not. test_kpi_source_note_is_exact_empty_string... is kept as-is. - attention_view._evasion_deltas duplicated evasion_deltas; it now delegates, with a test pinning that the two agree across every delta shape including a non-string kind. - Roadmap: record explicitly that the consensus-archive work must add a live reader of dis…2248b5b -
docs: prospection plan, tasks 12-14 and self-review
Demo builder, .eml export and CLI, playbook plus GDPR processing record, end-to-end suite asserting the pipeline runs with every paid provider off. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_016XzekHxqiXutBW4gw3Arsx00e69d8 -
docs: use the real agent/llm.py API in the compose task
call_claude does not exist; make_chat_model(default_config()) does. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_016XzekHxqiXutBW4gw3Arsxe7e6e9a -
docs: prospection plan, tasks 10-11
GDPR enforcement (hashed suppression, retention purge) and evidence-graded drafting with a fail-closed validator. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_016XzekHxqiXutBW4gw3Arsx32df131 -
docs: prospection plan, tasks 7-9
Paid provider adapters (Explorium MCP handoff, Dropcontact, LinkedIn), expiring intent signals, public holdings extraction with human confirmation. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_016XzekHxqiXutBW4gw3Arsx64054ca -
docs: prospection implementation plan, tasks 1-6
Store, config, AMF registry, entreprises API, two-pass scoring, enrichment waterfall. Tasks 7-14 follow. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_016XzekHxqiXutBW4gw3Arsx2369ff0 -
docs: spec the buy-side prospection sub-system
Design for prospection/: AMF open-data universe, deterministic ICP scoring in two passes, cost-ordered contact enrichment waterfall with all paid providers off by default, evidence-graded messaging, and a demo builder that refuses to ship a PARTIAL brief. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_016XzekHxqiXutBW4gw3Arsx31f1e63
-
fix: correct two false claims in the attention-ordered brief spec
Both were found while writing the implementation plan and both would have caused real breakage if followed as written. The spec instructed adding `accession` to verify_fact's locator comparison. That comparison (evidence.py:384-390) is strict equality returning FAILED, and the model does not reliably copy refs — evidence.py:365-372 already documents that it re-derives evidence_id rather than copying. A ref lacking the new field would have failed every fact in every brief. The field is now derived server-side in make_evidence_record, with a separate tolerant check, and is explicitly kept out of stable_evidence_id's identity dict so persisted evidence_ids stay valid. The spec also claimed tests/test_verdict.py already asserts brief block order. It does not — it holds pure-helper tests plus a NAV_ITEMS assertion. The plan extracts brief_block_order() so ordering becomes testable at all. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>6c906d3 -
docs: add implementation plan for the attention-ordered company view
Nine tasks, each TDD with a real failing test first. Sequence: traceability choke point (evidence_link + meta_row), accession as a first-class EvidenceRef field, KPI strip links, reusable top-N drift summary, a home for the homeless tone_trend signal, reusable evasion summary, the zone reorder itself, an i18n completeness guard, and hygiene. Two spec corrections found while reading the code: - The spec says tests/test_verdict.py "already asserts order". It does not -- it holds only pure-helper tests plus a NAV_ITEMS assertion. The plan therefore creates a render-order test, which is why render() is refactored to expose a pure brief_block_order() function. - The spec says to add `accession` to verify_fact's locator comparison. Doing that literally would fail every fact whose ref the synthesis model wrote without the new field -- the same hand-copying weakness the existing evidence_id fallback comment documents. The plan derives accession server-side in make_evidence_rec…f0edb96 -
fix: correct fetch_earnings return shape in the consensus archive plan
fetch_earnings returns (data, error), not a bare dict (alphavantage.py:68). The backfill task unpacked it as a dict and would have raised on .get(). Adds a self-review note recording the verified return shapes and the scripts/ package gap that put list_tickers() in storage/metrics_db.py. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>a00b388 -
docs: add implementation plan for the consensus archive
Seven tasks, TDD throughout. Task 1 builds the append-only JSONL store whose first test is the direct regression for the earnings_cache upsert that caused the missing history. Task 2 encodes ARCHITECTURE.md's alignment table as five independently tested refusal conditions. Task 7 replaces the hardcoded market_expectations_aligned=False with a computed value. The plan keeps analyst.py's own comparison_allowed=False deliberately: a single fetch cannot prove pre-release alignment, so that decision stays in one place. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>7968cf0 -
docs: add attention-ordered dashboard roadmap and first two sub-project specs
Restructures the product around one rule: what only this system can produce comes before what Bloomberg already shows. Decomposes the four-view target into seven sub-projects with an explicit dependency order. Three structural blockers found by code inspection and recorded in the roadmap: consensus is never archived (analyst.py:258 hardcodes comparison_allowed=False, so the scorecard block is dark by construction), no base rate exists (textdiff.py diffs one prior period against hardcoded thresholds), and the system emits no falsifiable prediction, so a calibration view has nothing to score. Two departures from the source target document, both argued in the roadmap: the radar and sector season are gated on reaching 30 covered names rather than sequenced by time, and calibration v0 scores the language signal instead of the model, because that is backtestable immediately. Sub-project B (consensus archive) starts first: its value accrues only with elapsed calendar time. Co-Authored-By:…d5b1cf3
-
docs: add implementation plan for The Brief dashboard redesign
7 TDD tasks covering the weighted grid, KPI sparklines, storyline compact mode, prose-tightening prompt rules, and a diagnosed fix for the "Since the last brief" same-quarter-rerun diff bug.acf6b3b -
docs: add design spec for The Brief dashboard redesign
Weighted grid layout, KPI sparklines, and AI-prose tightening to fix the flat single-column stack and generic text on the main brief screen.2775caa -
fix: restore evidence_snippet threading lost from an earlier uncommitted edit
The evidence_snippet fix (Ruling 1: thread history entries need the quoted text, not just evidence_ref's id/hash pointer) was applied to this file as an uncommitted edit before Task 1 was dispatched, then silently lost before being committed -- likely clobbered by a subagent's git operation in this shared worktree. Task 1's actual merged code (storage/threads_db.py, commit 0a8ebab) already has the fix correctly, confirming the loss was plan-document-only. Task 4's implementer caught the resulting signature mismatch and correctly reported NEEDS_CONTEXT rather than guessing. Re-applied and committing immediately this time. Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>9d243f9 -
fix: correct plan's continuity-category test to expect coercion
Task 2 review caught that the plan's own test test_brief_output_drops_invalid_continuity_thread_item contradicted the plan's own _normalize_continuity_category (which coerces unknown categories to Operational and keeps the item, matching CategorizedRisk's precedent) by asserting the item gets dropped instead. Split into two tests: one proving coercion-and-keep for a bad category label, one proving genuine structural invalidity (missing required field) is still dropped. Ruling recorded in .superpowers/sdd/2026-08-27-continuity-threads/progress.md. Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>e07ca52