docs: AI sales automation revamp implementation plan#1
Draft
gorajing wants to merge 2 commits into
Draft
Conversation
PLAN-anthropic.md sketches the strategic phases; the detailed plan under docs/superpowers/plans/ is the executable, TDD-style task-by-task version that supersedes the parent on conflicts. Iterated through 8 codex review rounds to clean: schema correctness (routing rules in Markdown only, partial unique indexes via the sqliteTable extra-config builder), idempotency at every entry point (signal ingest, scoring, routing, alerts via reserve-then-send), trustedSender gating for verification, deterministic timestamp ordering, and provenance preservation through connectors. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
gorajing
changed the title
gorajing
added a commit
that referenced
this pull request
May 10, 2026
Codex flagged 3 SHOULD-FIXes: 1. Per-rule rounding silently zeroed small weights for the last fraction of their window (a baseWeight=5 rule contributed 0 for ~10% of its window). Switched return type to fractional float; consumers in lib/scoring/score.ts will sum and round only at the final score. This also eliminates Math.round's sign asymmetry (+0.5 → +1 but -0.5 → -0) which was producing biased magnitudes for penalty rules. 2. baseWeight was documented as finite but not validated. NaN / ±Infinity propagating through the scoring sum would corrupt the final score for the entire account. Now throws TypeError before any math, mirroring the windowDays guard. 3. Negative-weight tests were inadequate (test value was already integral so rounding asymmetry was masked). Added a symmetric ±baseWeight test at quarter-window that asserts pos + neg ≈ 0 and |pos| ≈ |neg|, locking in the no-bias property. Plus three NIT fixes: - "+1ms in the future" comment misnamed (the date is BEFORE the event): renamed variable to justBeforeEvent and corrected the comment. - Added explicit -Infinity windowDays test (prior only covered +Infinity). - Added boundary test "1ms before full window" that asserts the float return is positive — locks in the bug fix from #1 above. Test count: 26 in this file (was 20), full suite 211/211 (was 205).
gorajing
added a commit
that referenced
this pull request
May 18, 2026
#1 Account identity: accept GitHub actor namespace (github.com/<actor>) for v1. Do NOT build an actor->CRM resolver now -- identity-from-actor is product logic, and a bad resolver is worse than none (contaminates account scores). v1.5: identity layer or operator-editable data/account-aliases.md. #3 GitHub all-or-nothing watch entries: accept fail-loud-per- connector for v1. 3.4's connector-level isolation means a dead repo no longer starves the other connectors and the failure is visible in the poll response. The half-fix (swallow per-entry failures) is explicitly rejected. No code change -- product calls. 3.2/3.3 NOT reworked.
gorajing
added a commit
that referenced
this pull request
May 18, 2026
…improves + nit) BLOCKER — wrong-message attribution via MUTABLE currentRevisionId (the user's #1 stated risk). Sequence: send rev1 → reply → redraft to rev2 → digest attributed rev1's reply to rev2's critique. engagement_events has no sent-revision marker. Fix: EXCLUDE a touch whose last touch_revision was created AFTER its first engagement event (ambiguous: which revision did the reply answer?). The normal pre-send flow (rev1 → critic_rewrite rev2 → SENT → events) has lastRev <= firstEvent and is correctly INCLUDED. Conservative + honest; v1.5 sentRevisionId removes the ambiguity. Test pins the redraft sequence is dropped from BOTH arms. WOULD-IMPROVES: - bounced touches excluded from the denominator: a hard bounce never landed, so counting it "silent" turned a deliverability failure into principle evidence. ('unsubscribed' = received, left as observed silent for v1.) Test pinned. - the mixed-timestamp test was decorative (2020 vs 2026 — naive local parse still picked right). Replaced with a TZ-INDEPENDENT direct parseTs unit test (exported parseTs) asserting SQLite space-format == ISO-Z == Date.UTC, plus a meaningful same-day integration test. Now genuinely pins the codex-r1 UTC blocker. - exported field names passed_*/failed_* invited the "pass" overclaim back → renamed to noFinding_*/flagged_* (+ flagged_lift). Module prose updated to no-finding/flagged; render labels already matched. NIT: parseTs('') → NaN is defensive (createdAt is NOT NULL from SQLite); covered by the direct parseTs test. Fixture correction: makeTouch now sets touch_revision.createdAt BEFORE its engagement events (10:00 < sent 11:00 < reply 12:00) — the redraft-ambiguity guard correctly exposed that the prior fixture (revision defaulting to ~now, events at 2026-05-06) was a temporally-impossible timeline. 580/580; typecheck + build + nightly-digest smoke clean.
gorajing
added a commit
that referenced
this pull request
May 18, 2026
Phase 6 is the proof/packaging phase, not a build phase. It splits
into agent-ownable (deterministic, safe) vs human-owned (judgment/
identity). This commit is the agent-ownable rails; the research,
prose, Loom, and submit remain human-owned ("do not let the system
outrank your judgment").
- PRIVACY GATE (constraint #1; also a real plan gap — Phase 0
never added it). .gitignore now ignores `application/*` (the
CONTENTS glob, not the directory) so the package is private by
default while keeping the plan's selective-publish escape hatch
(`!application/architecture-essay.md`) workable. evidence-pack /
outreach files stay private — they may name a hiring manager.
- lib/application/verify.ts — pure, fails-closed pre-submit gate.
Turns "every factual claim cites verified evidence" + "fact-check
prose before use" from a hope into enforcement: every `ev_…` id
cited in cover-letter.md must exist AND be `verified` in
evidence-pack.json, else the package is rejected. Also: required
files present, ≥1 outreach artifact, word count 500–800, and a
cover letter citing NO evidence fails (the proof claim itself
unbacked). Reports ALL problems at once. The evidence-id regex is
anchored to the exact newId('evidence') shape so prose can't
false-match (a loose pattern would launder an unbacked claim —
the precise failure this gate prevents). It does NOT judge prose
or whether to submit — mechanical floor only. 10 tests pin the
fail-closed behavior (a permissive verifier is worse than none).
- scripts/dump-evidence.ts — export VERIFIED-only evidence to the
gitignored application/evidence-pack.json. Fails loud on bad
usage (exit 2) / unknown account (exit 1, no garbage pack);
warns on 0 verified rows.
- scripts/verify-application.ts — thin IO shell over the verifier.
Exit 0 iff the package passes every mechanical check, 1 otherwise
with all problems listed. Gates the final
typecheck && test && build && verify-application step. Smokes:
no-arg→2, bad-account→1, no-package→1 (fails closed).
592/592; typecheck + build clean. NO live LLM / real-company action
taken — those are human-owned and delivered as a runbook.
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.
Adds the reviewed implementation plan for the AI sales automation revamp.
What this is
PLAN-ai-sales.md— strategic phases (high-level)docs/superpowers/plans/2026-05-06-ai-sales-automation-revamp.md— task-by-task TDD-style executable plan; supersedes the parent on conflictsReview boundary
This PR is for plan approval. Implementation lands separately from this plan branch and can be rebased onto
mainonce this lands.Plan stats
Out of scope here
🤖 Generated with Claude Code