2026-05-08

Focus

• Closing out CLI ergonomics and documentation onboarding for dont

Log

Session 15:55 (incitaciones)

• Explored .pi/agent/sessions/ to evaluate token consumption from tool loading.
• Found that Claude CLI already uses progressive disclosure: ~10 core tools loaded eagerly, ~40+ MCP/platform tools deferred (names only, schemas fetched on demand via ToolSearch).
• Identified actual token waste: Google suite (Gmail, Drive, Calendar) and Playwright MCP servers load globally for every project session even when irrelevant.
• Recommended moving those MCP configs from ~/.claude/plugins/ to per-project .claude/settings.json scope.
• Identified skill index bloat: ~80 skills listed per turn, with incitaciones:* namespace duplicating many global skills.
• Next: Audit MCP server placement and deduplicate incitaciones vs. global skills to shrink per-turn system-reminder size.

Session 16:58 (general)

• Extended the custom-provider-claude-cli pi extension to display Claude rate limit info in the pi statusline.
• Discovered the stream-json CLI format uses a separate rate_limit_event (not embedded in result); it carries status ("allowed"/"rejected") and resetsAt but no used_percentage.
• Debugged via event logging — synchronous readFileSync/writeFileSync on every event froze the pi UI; removed debug code.
• Final state: statusline shows quota: 5h: resets Xh Ym using resetsAt timestamp from rate_limit_event.
• File modified: ~/.pi/agent/extensions/custom-provider-claude-cli/index.ts
• Next: used_percentage is only available from pi's own status payload API — not accessible from a subprocess extension.

Sessions 12:45 (dont — 3 sessions)

• Implemented atom management subcommands (define, dismiss) with persistence and rendering; closed dont-72t.
• Completed and archived improve-cli-ergonomics OpenSpec (16/16 tasks); archived as 2026-05-08-improve-cli-ergonomics.
• Evaluated documentation quality (UX/DX diagnostic + Anti-slop audit); confirmed clean prose but identified tutorial gap.
• Enhanced mdBook: created "Getting Started" tutorial (tutorial.md) and refactored guides for Diataxis modality alignment.
• Code quality: Resolved clippy::collapsible-if warnings and syntax errors in src/main.rs.
• Arc: Moved from feature completion to UX/DX refinement, bridging the gap between implementation and user onboarding.
• Next: dont-3so (validate ergonomics milestone) and work through remaining 5 ready issues (shell completions, config parser, layout validation).

Session 20:02 (atril)

• UI alignment session (/ui-align): identified three pain points — no in-app back trail, no browseable entry point, dead-end cross-link jumps.
• Spec agreed: B→A+C mental model (category cards → file tree + search), breadcrumbs, persistent nav, back stack.
• Issue review (/issue-review): found 4 existing tickets (atril-6v4, atril-94t, atril-r2n, atril-xgv); removed artificial dep atril-94t→atril-xgv; added 2 missing tickets (atril-0mo slide-in file tree, atril-0jh persistent search).
• Implemented atril-6v4: replaced navDepth integer with history.state-backed NavState { depth, view, previousView }; contextual back labels via VIEW_BACK_LABELS; reload-safe (routeTo on reload, not navigate); shared-link first-visit safety (falls back to navigate(overview) instead of ejecting user).
• Ro5 review found 3 real issues: EDGE-001 (eject on shared link), CORR-001 (navDepth redundant parallel state), CORR-002 (unsafe cast). All three fixed.
• Playwright MCP validation confirmed: back nav, contextual labels, reload-then-back all working.
• Fixed pre-existing CI failure in navigation-renderers.ts (TS2488 on double-destructure).
• atril-6v4 closed; commit pushed.
• Next: atril-r2n (breadcrumb) or atril-xgv (persistent nav) — both unblocked; state shape { depth, view, previousView } ready to consume.

Sessions 17:50 (dont — 6 sessions)

• Layout validation (dont-8os): Project::open() now validates .dont/ subdirs with LayoutInvalid error, exit code 3, and dont init remediation; Ro5 review + 3 fixes; pushed.
• Config parser (dont-2tl, dont-h8c): new src/config.rs with structs for [import], [verify_evidence], [define.shape], [trust.hedges], [rules]; wired into dont import, verify-evidence, validate_label, dont trust, dont ignore, dont define; pushed.
• Rule engine (dont-8cs): RuleEngine in src/rules/mod.rs with evaluate(), severity(), warn/strict dispatch, mode escalation; Store::run_rule_query() wired; 9 tests.
• 7 shipped rules (dont-711, dont-7bg, dont-v70, dont-5sq): ungrounded, unresolved-terms, stale-cascade, lockable, correlated-error, dangling-definition, term-nonfunctional-label; Ro5 review → extracted source_key/host_from_uri, malformed-dep guard; 46 tests.
• Rules CLI (dont-y8v, dont-iny): dont rules list/show/add/test with 18 integration tests; SHIPPED_RULES registry; severity display.
• /rule-of-5-universal review → 10 findings fixed: shipped-name shadow guard (cannot-shadow-shipped-rule), compile-error remediation hint corrected, rows_to_matches now propagates type errors, --force required to overwrite existing rules, TOCTOU removed from Test handler, severity_label → &'static str, load_config() hoisted, strict-over-warn comment; 3 new tests (21 rules tests total).
• just install succeeded; binary updated.
• Arc: Full rule engine pipeline shipped end-to-end: engine → 7 rules → CLI subcommands → quality hardened by two Ro5 passes.
• Next: dont-9fv (dont explain command, P2), dont-8h2 (shell completions, P2), or dont-fl6 (transaction ID sequencing bug, P2).

Sessions 21:10 (dont — 3 sessions)

• Enforcement gate (18:20 session): investigated how dont enforces claims without a bypass path — enforcement is in-binary (Rust match + exit 1). Added dont prime as a pre-commit gate in prek.toml; wrote docs/enforcement.md covering all 6 gates and git/Claude Code/CI integration; applied 5 Ro5 corrections.
• rule-claim-structure lint rule (20:17 session): implemented full Phase 2 of add-rule-claim-schema OpenSpec — Rust rule, config, registration, translation doc, howto; 13 Ro5 issues found and fixed (doubted claims skipped, 2 new tests, docs for edge cases and troubleshooting); OpenSpec archived as 2026-05-08-add-rule-claim-schema; pushed at 78f60f4.
• dont-fl6 (tx ID sequencing bug): replaced next_tx() max(tx) query with a tx.seq counter file read/written under the existing with_write_lock(); migration seeds from max(tx) on first use; added two tests (8 concurrent Store threads, 8 parallel dont conclude subprocesses); closed and pushed.
• dont-dmk (drifted evidence count): added drifted_evidence key to dont prime --json assessment_counts; counts claims and terms with at least one repo-file evidence entry where fingerprint has drifted; TDD (test first); closed and pushed.
• Arc: Bug hardening and observability — tx sequencing is now a structural guarantee (not SQLite emergent), and prime surfaces drift counts for CI/gate use.
• Next: dont-9fv (dont explain command, P2) or dont-8h2 (shell completions, P2) — both unblocked.

Sessions 21:11 (atril — 2 sessions)

• Implemented atril-94t: overview redesigned as 2×2 category cards (Specs, Docs, Issues, History); inactive cards grayed; active cards navigate with correct "← Overview" back labels; Playwright-validated
• Closed atril-94t; atril-r2n (breadcrumb) and atril-xgv (persistent nav) remain unblocked
• Arc: Navigation foundation complete — history.state back stack (atril-6v4) + category-card entry point (atril-94t) now in place
• Next: atril-r2n (cross-screen breadcrumb) or atril-xgv (persistent category nav bar)

Session 20:52 (general)

• Analyzed ~/.pi/agent/sessions/ (340 sessions, 103 MB) and ~/.claude/sessions/ to identify token anti-patterns.
• Key findings: 45.6% of sessions hit compaction; sequential read-modify-test loops without truncation; API retry noise (Gemini 429 retried 5× in-context); cached spec files consuming most of the context window; multi-hour monolithic sessions.
• Added Context discipline section to ~/.claude/CLAUDE.md and ~/AGENTS.md via chezmoi; pushed dotfiles.
• Fixed custom-provider-claude-cli extension: corrected contextWindow from 1,000,000 → 200,000 for Sonnet/Opus 4.6; added truncateContent() (4000-char cap per message) and 20-message sliding window to format-context.ts.
• Tracked both pi extensions (custom-provider-claude-cli, custom-provider-gemini-cli) in chezmoi under dot_pi/agent/extensions/; pushed dotfiles.
• Next: Audit MCP server placement (move Google/Playwright to per-project scope) and deduplicate incitaciones vs. global skills to shrink per-turn system-reminder size.