EPIC cas-9508: Factory session friction + build infra (7/11 closed) by pippenz · Pull Request #3 · pippenz/cas

pippenz · 2026-04-23T15:55:15Z

Summary

Seven tasks from EPIC cas-9508 (Factory session friction) plus build tooling & docs. Based on a P1 friction report from yesterday's abundant-mines session; addresses the supervisor/worker pain points that added ~40 min of overhead per multi-worker EPIC.

Closed in this batch (7/11):

cas-5572 — action=mine race: worker identity matching widened to {agent_id, agent_name, CAS_AGENT_NAME, CAS_SESSION_ID} with case/trim tolerance. Eliminates the spawn-race kick pattern.
cas-4181 — Factory TUI epic hijack: unified init + runtime picker via shared helper; strict-improvement gate on EpicStarted; ghost-epic liveness escape. Compile-time enforced.
cas-1a7c — Lease/status divergence: action=release auto-recovers lease-less InProgress tasks with audit note; new action=reset verb for dead-session recovery.
cas-3086 — Close flow: persist worker ReviewOutcome envelope on task record at close attempt; supervisor-close forwards persisted envelope without bypass_code_review; epic-close short-circuits when all subtasks carry valid receipts (with P0 defense-in-depth so subtask residuals still block).
cas-a9ab — SessionEnd hygiene: manifest written to .cas/logs/factory-session-{date}.log; gc_report MCP verb; git rev-parse --git-common-dir for correct main-worktree resolution inside factory layout.
cas-aeec — SessionStart WIP triage: supervisor-only banner surfaces uncommitted state with per-file (last touched by cas-xxxx) attribution; 20-row cap with gc_report overflow.
cas-7f57 — Director dup messages: TeamsManager::write_to_inbox dedupes identical (from, text) within 10-min window + prunes read messages >2h. Unread messages preserved regardless of age (supervisor recovery prompts cannot evaporate). All ops under flock(LOCK_EX).

Side deliverables:

.cargo/config.toml — linker swapped from lld to mold (typically 2-5× faster link, dominates incremental Rust test loop). Paired with cargo-nextest installed host-wide.
docs/requests/ — session observation report + 4 cross-team request docs from Petra Stella Cloud.

Still open under EPIC cas-9508 (for follow-up session): cas-4513 (Claude Code JS-crash detection, 514-line WIP salvaged), cas-d0f9 (supervisor-skill preflight, doc-only), cas-2749 (worker crash recovery), cas-6cb0 (typed LeaseNotFound, P3).

Every closed task has verifier approval receipts in its close reason + ReviewOutcome envelopes persisted on the task record.

Test plan

cargo nextest run -p cas-cli green (covers mcp_tools_test, hooks::, session_hygiene, factory suite — all confirmed green at close time on individual worker branches; merge-combined run needs verification)
cargo build --release clean
Factory TUI smoke test: start session, confirm epic panel tracks active epic (not lex-greatest)
Fresh worker spawn + action=mine returns assignment within 1s without coordination kick
Supervisor-close on a worker-jailed task forwards persisted ReviewOutcome without bypass_code_review=true

🤖 Generated with Claude Code

Structured 7-item observation log from abundant-mines EPIC run (supervisor quick-sparrow-64, session 48456bb1). Covers spawn-time action=mine race, verification-jail cost under outdated binary, orphan untracked files, stale InProgress leases, director dup messages, and bypass_code_review UX gap. ~40 min of supervisor friction quantified for triage.

mold is a drop-in replacement for lld that is typically 2-5× faster on the link step, which dominates Rust's edit-test loop wall time on incremental builds. Keeps linker=gcc; only changes -fuse-ld arg. Requires: mold package installed (apt install mold). Companion: cargo-nextest installed globally for faster test execution (cargo install cargo-nextest --locked). No config file change needed for nextest; use `cargo nextest run` in place of `cargo test`.

Init (`detect_epic_state`) and runtime (`DirectorEventDetector::detect_changes`) used different tiebreakers when multiple Open-with-branch epics existed: init picked the one with most active subtasks; runtime picked the lexicographically greatest ID. A stray later-ID Open-with-branch epic appearing mid-session would fire `EpicStarted` and overwrite `epic_state`, which then scoped `filter_director_agents_to_current_session` to the wrong epic and blanked the Tasks panel. - Extract the shared picker `pick_best_open_branch_epic` + score helper `open_branch_epic_score` in `director::events` and call it from both paths. - Change `detect_changes` signature to take `current_epic_id: Option<&str>` and gate Open-with-branch `EpicStarted` on strict improvement: a fresh zero-subtask epic cannot hijack an already-tracked active epic. - Two regression tests in app::tests cover the hijack scenario and the init-equivalent runtime pick. cas-4181

Review-time follow-ups for cas-4181: - Strict-improvement gate now treats `current_epic_id` as `None` if the tracked epic is no longer present in `data.epic_tasks` (closed/deleted between refreshes). Prevents the TUI from freezing on a ghost id when a legitimate new Open-with-branch epic appears. - Regression test added. - Fix stale `detect_changes_for_epic` name in picker doc comment. cas-4181

Session-end hygiene work in flight when worker harness died. Modified: handlers.rs, handlers_session.rs, factory_ops.rs, cas-supervisor-checklist builtin. New: session_hygiene.rs. Not verified — downstream worker should review before building on. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

…ests Review-time follow-ups for cas-a9ab: - `main_worktree_path` now asks git for `--git-common-dir` when cas_root sits inside a linked worktree. Without this, the factory layout `<repo>/.cas/worktrees/<name>/.cas` caused the manifest to label the worker's own WIP as "main worktree", inverting the supervisor triage promise. Falls back to `cas_root.parent()` if git is unavailable. - Added four tests (all green): - `porcelain_entry_label_covers_documented_codes` pins the six label arms + documents current fall-through for rename/unmerged. - `manifest_is_append_only_across_multiple_sessions_same_day` guards the multi-session breadcrumb contract. - `manifest_records_clean_tree_marker` asserts the `(clean)` line. - `main_worktree_path_resolves_to_main_repo_from_linked_worktree` covers the factory worktree layout regression. cas-a9ab

… (cas-aeec) Supervisor SessionStart now prepends a "⚠ Prior-factory WIP detected" banner to the hook context when the main worktree is dirty. Each entry carries per-file attribution derived from `git log -1` of the last commit that touched the file — the first `cas-xxxx` task id in that commit's subject+body is surfaced alongside the porcelain status. Untracked files render as "unattributed — no git history"; commits without a task id render as "no task id in last commit". The banner never mutates anything. New module pieces in session_hygiene: - `extract_task_id`: canonical 4-hex cas-id parser (case-insensitive, rejects non-hex, over-long runs, and adjacent-alnum non-boundaries). - `attribute_file_to_task`: `git log -1` wrapper returning the first cas-id from the latest commit touching a path. - `build_session_start_wip_banner`: composes the full banner or returns None when clean / git unavailable / cas_root unresolvable. SessionStart wires the banner into supervisor-role context (prepended so it hits the preview window). Non-supervisor sessions and clean trees fall through silently. cas-supervisor-checklist step 6 updated to document the banner and point at gc_report for on-demand re-run. Tests (3 new, 11/11 session_hygiene green; 210/210 hooks:: green): - extract_task_id_accepts_canonical_and_rejects_garbage - attribute_file_to_task_finds_cas_id_in_last_commit - build_session_start_wip_banner_renders_attribution_and_skips_clean cas-aeec

…s-aeec) Review-time follow-ups for cas-aeec: - P1 adversarial: Unbounded git log fan-out. Each attributed row spawns `git log -1`; a pathological dirty tree (hundreds of files after a prior-session crash) turned SessionStart into a multi-second stall that the user reads as "Claude Code hanging on startup". Cap banner at 20 inline rows via WIP_BANNER_MAX_ENTRIES; overflow prints "... and N more — run gc_report for the full list". - P2 adversarial: Banner prepended above codemap / project-overview, displacing them from the SessionStart preview top slot the adjacent comments explicitly engineered those modules to land in. Append instead (project-overview already does the same for the same reason). - P3 correctness: Widen label column from {:10} to {:15} so "modified-staged" (15 chars) no longer overflows and misaligns the attribution column. One regression test added (build_session_start_wip_banner_caps_rows_on_large_wip_trees) plus adjusted inner-loop width. 12/12 session_hygiene green. cas-aeec

Dropped by cloud team 2026-04-23. Two features (cloud sync pull team-memories for active project; make team config global with per-project override) + two bugs (push client must detect server-side skipped rows — complement to cas-0bdc; cloud client 404 on session start follow-up against existing cas-4244). Task IDs referenced by cloud team (cas-e38e, cas-eb38) do not yet exist in this DB — need to be created as cross-team tracking tasks in a future session. cas-f645 is already open (EPIC cas-2eb3); cas-4244 was closed via feature-flag workaround, re-filed doc asks for real route fix.

Partial director dup-message work in progress when session ended. Modified: cas-cli/src/ui/factory/daemon/runtime/teams.rs. Not verified — downstream worker should review before building on. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

…f57) Review-time follow-ups for cas-7f57: - Adversarial P1: retention was age-only, so an unread supervisor recovery message to a wedged worker would silently evaporate after 2h on any subsequent inbox write. `retain` now preserves `read: false` messages regardless of age; only already-delivered messages are subject to the 2h sweep. - Efficiency (maintainability P3 / adversarial P3): the dedup early-return used to re-serialize + write the inbox file even when retention had pruned nothing, producing a write storm under hot duplicate senders. Track `retention_pruned` and only persist on the skip path when something was actually removed. - Two regression tests added: - write_to_inbox_does_not_dedup_past_window (15-min-old duplicate passes through; guards dedup-cutoff sign flip) - write_to_inbox_retention_preserves_unread_messages (3h-old unread supervisor recovery message survives the sweep; 3h-old read nag is still pruned) - Existing retention test adjusted (stale message marked read=true to match the post-fix policy). 13/13 teams tests green. cas-7f57

… (cas-5572) Factory workers assigned by supervisor via `task update assignee=<worker-name>` were reporting "No open tasks" on their first `action=mine` poll, even though `task show` confirmed the write. Root cause: `cas_tasks_mine` matched assignee strictly against `agent_id` or the agent-store row's `name` — at spawn time the row may still carry a stale default name, so the filter missed name-based assignments. Widen the filter to consider a de-duplicated identity set built from `agent_id`, the agent-store name, `CAS_AGENT_NAME`, and `CAS_SESSION_ID` (the two env vars populated by the factory PTY spawner). Compare case-insensitive on trimmed values so minor drift (whitespace, casing) doesn't hide assignments either. Adds two regression tests in mcp_tools_test/task_tools that reproduce the reported scenario and lock in the trimmed / case-insensitive behavior. EPIC: cas-9508 — Factory session friction Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

…erb (cas-1a7c) When a factory worker session dies mid-task, the lease disappears from the agent store but the task row stays InProgress with the dead worker's assignee. `action=release` used to error with "No active lease found" and leave the status divergence in place, forcing operators to manually run `update status=open assignee=…` to recover. Two fixes for EPIC cas-9508 factory-session friction: 1. `cas_task_release` now detects the lease-less, non-Open, non-Closed case and auto-recovers: transitions to Open, clears assignee, writes an audit note to task.notes, and returns a success message flagging the recovery. The real "nothing to release on an Open task" path still errors. 2. New `action=reset` verb — an atomic "revive this task from a dead session" operation. Force-releases any active lease on the task (no ownership check), clears assignee, forces status=Open, writes an audit note. Refuses Closed tasks (tells the caller to use `reopen` instead). Also adds a regression test for AC3 confirming `action=show` immediately after `action=update` reflects the new status — no read-after-write snapshot lag in the task-store path. 5 new tests; full mcp_tools_test suite passes (132/132). EPIC: cas-9508 — Factory session friction Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

Adds four new close_ops unit tests: - persisted_envelope_satisfies_gate_when_req_missing - persisted_envelope_with_p0_still_blocks - request_envelope_takes_precedence_over_persisted - persisted_malformed_envelope_is_rejected Plus a full-stack integration test in verification_flow.rs (test_close_forwards_persisted_review_envelope_after_jail) that walks the cas-3086 scenario end-to-end: worker close with clean envelope hits verification jail, envelope is persisted, supervisor close without bypass_code_review=true succeeds by forwarding the persisted receipt.

Verifier feedback: epic_subtask_receipts_cover accepted any structurally valid envelope, so a subtask with a P0 residual receipt (reachable via direct DB write or a leak past the subtask close gate) would have let the epic-close skip the multi-persona gate. Defense-in-depth: require evaluate_gate(residual) == Allow for every subtask envelope before treating the union diff as pre-reviewed.

Includes: cas-4181 (TUI epic hijack fix), cas-a9ab (SessionEnd hygiene + gc_report), cas-aeec (SessionStart WIP triage banner), cas-7f57 (director dedup + retention).

P2 #4 — show_help was missing from alt_screen_scroll_input guard. When the help overlay is open and the focused pane is in alt-screen, wheel events were incorrectly forwarded to the inner PTY. Fix: add `|| self.show_help` to the suppression guard and update the doc comment to enumerate all overlay conditions. P2 #5 — PgUp (ESC [ 5 ~) and PgDn (ESC [ 6 ~) were not forwarded in alt-screen mode; the generic ESC handler only consumed the first byte (bare 0x1b) and left the remaining [5~ as stray ASCII. Fix: add a dedicated 4-byte sequence check in client_input.rs before the arrow-key check. In alt-screen mode, the full native sequence is forwarded via mux.send_input; outside alt-screen, handle_scroll_up/down is called (same behavior as mouse wheel). Satisfies AC #1 ("same on PgUp"). Added FactoryApp::for_test() — #[cfg(test)] minimal constructor that builds a FactoryApp without spawning PTYs or opening files, enabling unit tests for guard logic at the FactoryApp layer. Tests added: - sidecar_and_selection::tests (11 new unit tests): - Baseline: returns Some when guard clear + pane in alt-screen - show_help guard blocks forwarding (P2 #4) - MC + mc_focus==None guard blocks forwarding (P1 #1 regression) - MC + mc_focus==Workers guard blocks forwarding - PgUp/PgDn dispatch fires when alt-screen active (P2 #5) - PgUp/PgDn fall through to handle_scroll when NOT in alt-screen - Wheel scroll no-regress when not in alt-screen (AC #3) Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

Per supervisor amendment to cas-ec8f AC: fix the pre-existing test failure surfaced by my discovery note. The fix turned out to be two loops (not one), both stale from the v2.13.0 supervisor-default flip: - Loop 1 (cas-worker.md markers, line 952): previously asserted both "cas-code-review" and "close-gate.md". Commit 8b82273 / cas-8962 ("remove pre-close /cas-code-review step from worker skill") deliberately removed the cas-code-review mention from cas-worker.md when owner=supervisor became the default. Drop the now-stale "cas-code-review" assertion; keep "close-gate.md" — that's still the correct pointer from SKILL.md to the gate doc. - Loop 2 (close-gate.md content markers, line 959): previously asserted five inline-block markers — "Close-time Code Review Gate" (old section title), "If close is blocked on P0" (legacy P0 hard-block), "bypass_code_review" (legacy worker bypass), plus "cas-code-review" and "code-reviewer". Commit 167c57e ("docs(skills): finish cas-5815 supervisor-default flip — purge stale worker-runs-review prompts") rewrote close-gate.md to encode the supervisor-owned contract — the inline-block markers no longer apply. Replace with three markers that encode the *current* contract: "Close Gate" (section title), "cas-code-review" (skill still mentioned, but in a 'don't invoke pre-close' caveat), and the literal `owner = "supervisor"` flag. Both inline comments now record the historical context so the next person who deletes a marker from these docs is forced to notice the test pin and re-establish a meaningful contract instead of slipping into another silent-failure state. Verified: `cargo test test_cas_worker_skill_documents_code_review_gate` passes (was failing on main since 8b82273); the two new cas-ec8f regression tests still pass; cas-f645's push_skipped integration tests still pass. This is commit #3 on factory/daring-swan-93 (dc95458 cas-f645 + 16f5ba8 cas-ec8f Round 1 + this), per supervisor instruction not to rebase the prior two.

…screen cas-72c3 was the deferred test rider for cas-d5fa, blocked on cas-11b0 (FactoryApp::for_test constructor). Per cas-11b0's close note, the for_test() constructor + 8 unit tests at the FactoryApp layer already shipped in 53a7bf4 (same day cas-d5fa landed), covering: - AC #1 (MC + mc_focus==None + alt-screen → no forward): scroll_blocked_by_mc_focus_none - AC #2 (show_help + alt-screen → no forward): scroll_blocked_by_show_help - AC #4 (non-alt-screen no-regress): wheel_scroll_no_regress_when_not_in_alt_screen The genuinely uncovered slice was AC #3 — the daemon-dispatch path in `client_input.rs` lines 157-187 that maps `MouseScrollUp` to either `diff_scroll_up()`, `mux.send_input(SCROLL_UP_ARROWS)`, or a no-op, depending on `show_changes_dialog` and `handle_scroll_up()`'s return. That branch lives inside a deep `tokio::select!` and can't be exercised directly from a unit test without spinning up the full client loop, so this commit pins the three contracts the daemon relies on instead: 1. scroll_arrow_consts_have_exact_byte_shape_cas_72c3 Asserts SCROLL_UP_ARROWS == `ESC[A` × SCROLL_LINES and SCROLL_DOWN_ARROWS == `ESC[B` × SCROLL_LINES. Pre-cas-72c3 the only structural assertion (line ~37) checked length, not bytes; a typo could silently break wheel forwarding without firing any guard. 2. scroll_changes_dialog_blocks_alt_screen_forwarding_cas_72c3 With both show_changes_dialog AND alt-screen active, `handle_scroll_up` must return `Done` (not `AltScreen`). That's the post-condition the daemon's outer `if self.app.show_changes_dialog` branch relies on to early-return without forwarding arrows to the PTY. Pins both up and down directions. 3. daemon_dispatch_table_for_mouse_scroll_up_cas_72c3 Table-driven mirror of the daemon's three-way decision (`diff`, `alt`, `noop`), asserted against five FactoryApp state shapes. Any future drift in either the daemon dispatch or `handle_scroll_up`'s return contract fails the test row that no longer maps correctly. All 11 sidecar_and_selection tests (8 pre-existing + 3 new) pass. Test-only diff; no production code touched. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

EPIC cas-ffc4 child task cas-53d5 closes hypothesis #2 from the original cloud-team bug doc: `CloudSyncer::pull_team` now keys its `since=` watermark by `(team_id, project_canonical_id)` instead of by `team_id` alone. This prevents a user working on team T across two projects P1 and P2 from seeing their second-project pull silently skip historical backfill (same "0 of everything" symptom cas-6ec7 fixed at the endpoint-routing level). `pull_team` signature takes `project_id: &str` as an explicit parameter; `cas cloud pull --full` now scopes its watermark clear to the current (team, project) pair. EPIC cas-ffc4 remains OPEN — sibling child task cas-1ced (eager slug resolution at `cas cloud team set` to close hypothesis #3) will ship as a follow-on patch. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

Before this fix, `cas cloud team set <uuid>` printed "Slug resolution deferred — see `cas cloud team show`" and did NOT eagerly resolve the project canonical id. A new team member who cloned the repo into a directory whose basename didn't match the canonical slug (daniel.l: clone named `cas` vs canonical `cas-src`) had their first `cas cloud sync` go out with `project_id=cas` — silently routing push/pull to a phantom project. Manual workaround was a directory rename. EPIC cas-ffc4 hypothesis #3, the last lurking onboarding regression. Changes: - `config::ProjectConfig`: new struct with `canonical_id: Option<String>`, surfaced as `[project]` in `.cas/config.toml`. Wired into the main `Config` struct + `merge_missing` so it round-trips through the existing TOML loader without touching any sync paths. - `cloud::config::resolve_canonical_id`: prepended a config.toml read to the resolution chain (config.toml → folder-name → path-hash). Fully backward-compatible — when no [project] block exists, the chain is exactly the prior behavior. New helper `canonical_id_from_config_toml` is the read side; `set_canonical_id_in_config_toml` is the write side with read-modify-write semantics that preserve other sections (verified by `config_toml_preserves_other_sections`). - `cloud::config::derive_canonical_id_from_git_remote` + `normalize_git_remote_url`: derive from `git -C <root> remote get-url origin`, normalize to `<host>/<owner>/<repo>` form. Handles HTTPS, HTTP, SSH `git@host:owner/repo`, and `ssh://git@host/owner/repo` shapes; strips `.git` suffix; rejects unrecognizable shapes (local paths, etc.) so the caller falls through to the next resolution step. - `cli/cloud.rs::execute_team_set`: after persisting team_id, runs the resolution flow (config.toml → git remote → defer). Source of resolution surfaced to both human and JSON output. The "deferred" message survives ONLY when neither config nor git yields a slug — explicitly does NOT default to the working-dir basename (the bug this task fixes). Signature now takes `cas_root: &Path` for test isolation. - New `cas cloud project set <canonical-id>` subcommand: manual override path for monorepo / custom-layout / non-git directories where auto-derivation fails. Writes `[project] canonical_id` to config.toml. - `cli/cloud.rs::execute_team_show`: now displays the resolved project slug alongside team UUID + team slug. JSON output adds `canonical_id` key. - `execute_team_show_for_test` helper exported for integration tests (returns the JSON-string payload directly so tests don't need stdout capture). Tests: - 6 new integration tests in `team_set_slug_resolution_test.rs`: config.toml preserve / HTTPS-derive / SSH-derive / no-default-to-basename negative / project set writes / team show displays. - 11 new unit tests in `cloud::config::tests`: URL normalization shape table (https/http/ssh/ssh-url/subgroup/local-rejected/empty-rejected), config.toml round-trip, preserves-other-sections, resolution-order precedence. 29 cloud-sync integration tests + 32 cloud::config unit tests + full crate suite all green. Build clean. Onboarding-only surface — no production sync path touched; the prepend to `resolve_canonical_id` is fully backward-compatible when no [project] block exists. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

EPIC cas-ffc4 child task cas-1ced closes hypothesis #3 from the original cloud-team bug doc: `cas cloud team set <uuid>` previously deferred slug resolution entirely, causing the first sync to go out with the wrong `project_id` when the working-directory name didn't match the canonical slug (daniel.l's `cas` vs canonical `cas-src`). Manual directory-rename was the workaround. Fix: - New `[project]` config block (`ProjectConfig { canonical_id }`) wired through Config + merge_missing + init.rs literal. - `resolve_canonical_id` chain becomes config.toml → folder-name → path-hash. Backward-compatible. - New helpers in `cloud::config`: `canonical_id_from_config_toml`, `set_canonical_id_in_config_toml` (read-modify-write preserves other sections), `derive_canonical_id_from_git_remote`, `normalize_git_remote_url` (HTTPS / HTTP / `git@host:owner/repo` / `ssh://git@host/...` / `.git` suffix strip). - `execute_team_set` runs eager resolution after team_id persists; output surfaces source (`from .cas/config.toml` / `derived from git remote` / deferred). Explicitly does NOT default to working- dir basename — that's the exact bug guard. - New `cas cloud project set <canonical-id>` subcommand for manual override (monorepo / non-git / custom layout). - `cas cloud team show` displays the resolved project slug. Tests: 6 new integration tests in `team_set_slug_resolution_test.rs` (preserve config / HTTPS derive / SSH derive / no-basename-default negative / project set / team show) + 11 unit tests in `cloud::config::tests` (URL normalization shape table + config.toml round-trip + section-preserve + resolution precedence). This is the last task in EPIC cas-ffc4 — the EPIC closes on merge. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

…loses EPIC cas-ffc4 closes end-to-end. Last child task cas-1ced closes hypothesis #3 from the cloud-team bug doc: `cas cloud team set <uuid>` now eagerly resolves the canonical project slug via .cas/config.toml → git remote → defer (NEVER falls back to working- dir basename). New `cas cloud project set <canonical-id>` subcommand for manual override. `cas cloud team show` displays the resolved slug. This was the last UX paper-cut from daniel.l's original bug report — the directory-rename workaround is no longer needed. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

…c4 closes Closes the original cloud-team P1 end-to-end. EPIC cas-ffc4 had three child tasks mapping to the three hypotheses from the bug doc: - cas-6ec7 (v2.15.1): hypothesis #1 — missing endpoint wire-up in `cas cloud sync` / `cas cloud pull` - cas-53d5 (v2.15.2): hypothesis #2 — cross-project watermark reuse in `CloudSyncer::pull_team` - cas-1ced (v2.15.3, this release): hypothesis #3 — deferred slug resolution causing wrong `project_id` on first sync after `cas cloud team set` `cas cloud team set <uuid>` now eagerly resolves the canonical project slug from `.cas/config.toml` then git remote, surfaces the source, and explicitly does NOT default to the working-dir basename (that was the bug). New `cas cloud project set <canonical-id>` subcommand for manual override. `cas cloud team show` displays the resolved slug. Bug docs moved to docs/requests/completed/. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

pippenz and others added 16 commits April 23, 2026 10:36

Merge cas-9508 Lane B + TUI hijack work from factory/silent-owl-56

ea95adf

Includes: cas-4181 (TUI epic hijack fix), cas-a9ab (SessionEnd hygiene + gc_report), cas-aeec (SessionStart WIP triage banner), cas-7f57 (director dedup + retention).

pippenz merged commit f8f1381 into main Apr 23, 2026

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

EPIC cas-9508: Factory session friction + build infra (7/11 closed)#3

EPIC cas-9508: Factory session friction + build infra (7/11 closed)#3
pippenz merged 16 commits into
mainfrom
develop

pippenz commented Apr 23, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

1 participant

Conversation

pippenz commented Apr 23, 2026

Summary

Test plan

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

1 participant