v0.5.11: fix stale HTML report on re-push

Bug fix: findRunFile returned the oldest timestamped report.html instead of the latest. When report was run twice (e.g., after adding backend logs), push uploaded the stale first HTML while run.json had the current data.

Root cause: readdirSync().find() returns the first match. Timestamped files sort chronologically, so the first match is the oldest.

Fix: Changed to filter().sort() and return the last (latest) match.

Test: 5 new tests for findRunFile (333 total).

Fixes the L3c3ZDTrZt report where HTML showed 651 entries (app only) but JSON had 1000 entries (app + backend + step).

v0.5.5: log timeline pipeline + timestamp-based file naming

What's new

Log timeline pipeline

• Log parser (src/log-parser.ts): auto-extract timestamps from any common format (ISO 8601, Python comma-ms, logcat, space-separated), detect log levels
• Timeline in HTML reports: auto-discover .log files in run directory → parse → filter to run time window → correlate with steps → render chronological timeline with source badges and clickable citations to raw log lines
• Raw log viewer: collapsible sections with line numbers and anchor IDs for citation linking

Timestamp-based file naming

• All run files now get a compact timestamp prefix: 20260330T100001200Z-events.jsonl, 20260330T100005000Z-run.json, etc.
• run.meta.json is the only fixed-name file — serves as directory index storing all timestamped filenames
• findRunFile() resolver (src/run-files.ts): backwards-compatible lookup that checks fixed name first, then scans for timestamped variant

Stats

• 322 tests (18 new)
• 18 files changed, +725 / -37 lines

v0.5.2

Fix 15 design principle violations found by codex audit

Docs: update CLAUDE.md command list (8 not 6), fix README command table, event types, run.json example, architecture. Remove stale references to run/migrate commands and non-existent files.

Code: fix exit codes (no-subcommand=2, unverified=2), add input validation for all --run-dir inputs, remove raw ADB instructions from schema, fix error hint text.

v0.5.1

Bug fix: step outcome ignoring tier verification results

Step outcome was derived solely from step.end event outcome field (defaulting to fail when absent). Two-tier verification results were computed but never fed back into step outcome. Now: tier failures override step.end pass, and when step.end has no explicit outcome, tiers determine the result. 291 tests pass (3 new).

v0.5.0: Agent-Friendly Verification

What's new

Honest results — Reports no longer show misleading green PASS when nothing was verified. New unverified result when all automated checks have no evidence and all agent reviews are pending. Audit mode shows AUDIT badge instead of PASS.

Step recipes — record init --flow <yaml> now returns per-step event recipes telling agents exactly what to stream:

{"recipe": [{"id":"S1","events":["step.start","action","artifact (screenshot: step-S1.webp)","assert (text_visible: Counter: 2, milestone: counter)","agent-review (prompt_idx: 0, verdict: pass|fail)","step.end"]}]}

agent-review events — New event type lets agents resolve tier 2 prompts:

{"type":"agent-review","step_id":"S1","prompt_idx":0,"verdict":"pass","reason":"Home tab bar visible"}

Event gap warnings — record finish now warns when expected events are missing (e.g., step has expect but no assert was streamed).

Install / upgrade

npm install -g flow-walker-cli@0.5.0

288 tests, 18/18 eval gates.

v0.4.1

Bug fix

• Fix multi-step flows with judge blocks collapsing to 1 step — The flow-parser's !inJudge guard prevented subsequent step IDs from being recognized after a judge: block. Uses indentation-aware detection now.
• Filter empty prompt cards in reports
• Footer shows package version (0.4.1) not schema version (3.0.0)
• --version shows both: flow-walker 0.4.1 (schema 3.0.0)

Install / upgrade

npm install -g flow-walker-cli

v0.3.2 — v2-only, all v1 code removed

What's Changed

Breaking: v1 support completely removed. All agents must use v2 flows.

Removed

• parseFlowV1() parser and LEGACY_STEP_KEYS
• migrate command (v1→v2 migration)
• v1 types: Flow, FlowStep
• v1 reporter: generateReport(), buildHtml()
• v1 run-schema: RunResult, StepResult, validateRunResult()
• v1 yaml-writer: generateFlows(), toYaml(), writeFlows()

Enforced

• parseFlowFile() rejects non-v2 flows with clear error
• report command validates v2 run.json (requires mode, steps[].outcome, steps[].do)
• Agents must use the pipeline: record init → stream → finish → verify → report → push

Stats

• 253 tests passing
• ~1,300 lines of v1 code removed
• Schema version: 2.1.0

Full Changelog: v0.3.1...v0.3.2

v0.3.1 — v2 enforcement + desktop video fix

What's new

v2 Schema Enforcement

• report command now rejects non-v2 run.json — requires mode, steps[].outcome, and steps[].do
• Clear error message tells agents to run flow-walker verify first
• No more silent "undefined" in reports from hand-crafted data

Desktop Video Recording Fix

• Replaced screencapture -v (broken by TCC) with ffmpeg via Terminal.app
• Terminal.app has Screen Recording permission — bypasses TCC entirely
• Tested on macOS Tahoe M4: 1920x1080, 10fps, h264, ~150KB for 20s

Pipeline Documentation

• record schema documents the full 6-step mandatory pipeline
• verify schema says "REQUIRED after record finish"
• report schema says "Requires verify — rejects v1/non-v2 data"

Stats

• 321 tests, typecheck clean
• Synced from beastoin/autoloop Phase 10+11

v0.3.0 — Desktop Support + Verify Fixes

What's New

Phase 10 — Desktop Support

• agent-swift bridge: AgentType (flutter | swift), detectAgentType(), platform-specific exec/textPress/bringToForeground
• CLI flags: --agent flutter|swift and --agent-path for desktop testing
• Desktop video: screencapture -v for macOS recording
• Schema v2.1.0: Updated command schema with new flags

Phase 11 — Verify Correctness + Polish

• Audit mode fix: result now reflects actual step outcomes (was hardcoded to "pass")
• Outcome normalization: "skip" → "skipped", "partial" → "fail", unknown → "fail"
• Expectation checking: text_visible and interactive_count now check assert event passed field
• YAML multi-line scalars: folded (>) and literal (|) support in flow parser
• Event schema validation: warns on non-standard step.end outcome values

Stats

• 320 tests passing (37 new)
• 15 files changed, 627 insertions

Validated

• 6 desktop flows run by sora on macOS via agent-swift
• Navigation (6/6), Dashboard (3/6), Chat (5/5), Memories (5/6), Tasks (4/5), Settings (9/9)
• Remaining failures are agent-swift AX gaps, not flow-walker bugs

v0.2.1

What's new in v0.2.1

Agent-first v2 architecture (Phases 7-9)

• 3-phase recording: record init → record stream → record finish
• NDJSON event streaming: step.start, action, assert, artifact, step.end, run.start, run.end, note
• Verify command: strict, balanced, and audit modes with VerifyResult schema
• Snapshot & replay system: auto-save coordinates after successful runs, 1.85x speedup on replays
• verify: true step marking: agents control which steps always get verified

Report improvements

• Dark-themed v2 HTML reports with embedded base64 screenshots
• Duration display from real wall-clock timestamps
• Auto-detect screenshots by step ID
• Video embedding support
• Landing page v2 pass count fix

Other

• Walk command generates v2 flows
• Auto video recording in record init/finish (use --no-video to disable)
• Comprehensive README rewrite covering full v2 pipeline

283 tests passing. Zero external dependencies.

Full changelog: v0.2.0...v0.2.1