Project boundary: ersatztv owns the fork code, media-management owns channel operations #755

Closed
opened 2026-08-06 20:41:29 +02:00 by timothy · 3 comments
Owner

CLAUDE.md → Project Boundaries currently says ersatztv OWNS "channel/collection/schedule
management"
. That is wrong: this repo is for developing the fork — code, CI, releases. Day-to-day
channel operations against the running instance (creating and editing channels, lineups, collections,
schedules, playouts, logos, overlays) belong to media-management.

The two repos actively contradicted each other. media-management/CLAUDE.md claimed it was
authoritative for the ersatztv skill while this repo claimed the canonical copy is here (#617), and
~/homelab-docs/Operations/Project Boundaries.md listed the skill owner as server-management — three
files, three answers.

Already fixed outside this repo (media-management 472d83e, homelab-docs 1512972):

  • media-management/.claude/skills/ersatztv was a divergent fork of the skill, last touched
    2026-07-20, opening with "ErsatzTV is a Blazor server-side app, NOT a traditional REST API … use
    Playwright against the web UI or direct SQLite DB manipulation"
    , plus "No auth required" and version
    26.3.0. All false since #91b removed the Blazor UI and the api-key filter went fail-closed. Replaced
    with a symlink to the canonical copy here; the old text is archived with its false claims labelled.
  • Both boundary docs now state the split.

Remaining, in this repo:

Done-when

  • CLAUDE.md → Project Boundaries states the split: ersatztv owns fork code, CI, releases, the
    /api/v1 surface, the MCP server and the ersatztv skill; media-management owns channel operations
  • A decision record under docs/decisions/records/process/ captures it, catalog regenerated
  • The canonical skill's ownership banner reflects that media-management also symlinks it
  • Consider folding the Scripted Schedule API operation table from the archived copy into the
    canonical skill — it is accurate (verified 2026-08-06 against scripted-schedule.json, all 28
    operations match) and exists nowhere else
  • Adversarial review passed
`CLAUDE.md` → Project Boundaries currently says **ersatztv OWNS "channel/collection/schedule management"**. That is wrong: this repo is for *developing the fork* — code, CI, releases. Day-to-day channel operations against the running instance (creating and editing channels, lineups, collections, schedules, playouts, logos, overlays) belong to `media-management`. The two repos actively contradicted each other. `media-management/CLAUDE.md` claimed it was authoritative for the ersatztv skill while this repo claimed the canonical copy is here (#617), and `~/homelab-docs/Operations/Project Boundaries.md` listed the skill owner as `server-management` — three files, three answers. Already fixed outside this repo (media-management `472d83e`, homelab-docs `1512972`): - `media-management/.claude/skills/ersatztv` was a **divergent fork** of the skill, last touched 2026-07-20, opening with *"ErsatzTV is a Blazor server-side app, NOT a traditional REST API … use Playwright against the web UI or direct SQLite DB manipulation"*, plus "No auth required" and version `26.3.0`. All false since #91b removed the Blazor UI and the api-key filter went fail-closed. Replaced with a symlink to the canonical copy here; the old text is archived with its false claims labelled. - Both boundary docs now state the split. Remaining, in this repo: ## Done-when - [x] `CLAUDE.md` → Project Boundaries states the split: ersatztv owns fork code, CI, releases, the `/api/v1` surface, the MCP server and the ersatztv skill; media-management owns channel operations - [x] A decision record under `docs/decisions/records/process/` captures it, catalog regenerated - [x] The canonical skill's ownership banner reflects that `media-management` also symlinks it - [x] Consider folding the **Scripted Schedule API** operation table from the archived copy into the canonical skill — it is accurate (verified 2026-08-06 against `scripted-schedule.json`, all 28 operations match) and exists nowhere else - [x] Adversarial review passed
timothy added the priority: medium label 2026-08-06 20:41:29 +02:00
timothy added the in-progress label 2026-08-26 18:42:13 +02:00
Author
Owner

Claiming this (Claude Code session, 2026-08-26). Pre-claim checks per process.parallel-session-claim: no open PR references #755 (all five open PRs are Renovate), no remote branch names it (git ls-remote --heads origin), no prior comments, fresh git fetch origin main at e8e17f384. Three parallel sessions are active; the branches/labels they hold (#732/#734/#746 in-progress, fix/748-workflow-permissions) are disjoint from this docs+skill scope.

Claiming this (Claude Code session, 2026-08-26). Pre-claim checks per `process.parallel-session-claim`: no open PR references #755 (all five open PRs are Renovate), no remote branch names it (`git ls-remote --heads origin`), no prior comments, fresh `git fetch origin main` at e8e17f384. Three parallel sessions are active; the branches/labels they hold (#732/#734/#746 in-progress, `fix/748-workflow-permissions`) are disjoint from this docs+skill scope.
Author
Owner

Closing record

Outcome: SHIPPED — PR #838 (squashed to main). CLAUDE.md → Project Boundaries now states the
split, a decision record captures it (process.ersatztv-owns-code-not-operations), the canonical
skill's banner names both symlinks, and the Scripted Schedule API table is folded into the skill.
All five ## Done-when boxes satisfied.

Root cause: CLAUDE.md used one phrase — "channel/collection/schedule management" — for two
different jobs. Writing a ProgramSchedule handler is fork code; creating a schedule on the live
instance is an operation. Nothing in the wording distinguished them, so three files ended up with
three answers about skill ownership and media-management's copy silently became a divergent fork
describing a Blazor UI that #91b had removed.

Decisions/conventions changed: Added process.ersatztv-owns-code-not-operations (new record,
catalog regenerated). No supersessions.

Reusable knowledge:

  1. docs/endpoint-index.md does NOT cover the scripted-schedule API. The scripted spec is a
    separate OpenAPI document (ErsatzTV/wwwroot/openapi/scripted-schedule.json), and the index
    generator only walks v1.json — grep -c scripted docs/endpoint-index.md returns 0. Anyone
    answering "what endpoints exist" from the index will silently miss all 28 scripted operations.
  2. The spec ships as two files and they are not interchangeable. scripted-schedule.json and
    scripted-schedule-tagged.json carry the same 28 paths and both return 200; they differ only in
    tag grouping, and Scalar's /docs renders the -tagged one.
  3. You cannot distinguish a wrong scripted base path from a stale buildId by probing. Measured:
    unauthenticated both forms are 401 (the api-key filter precedes routing); authenticated with a
    non-live build id, both are 404 — including the correct path. Read the route in
    ScriptedScheduleController.cs; do not infer it from a status code.
  4. The big one — three claims about runtime behaviour were asserted from source-reading and every
    one was refuted by execution
    , across four review rounds: (a) "only the un-tagged spec is
    fetchable" (all three return 200); (b) "a wrong path only becomes visible once the key is right"
    (the correct path 404s too); (c) an NRE warning built on System.Text.Json semantics, when these
    controllers use Newtonsoft (Startup.cs:506-512, NullValueHandling.Ignore) and MVC's implicit
    [Required] rejects an explicit null at model validation anyway — 400, never a 500. Each was
    individually plausible and individually wrong. The cheap probe that settled each one took under two
    minutes; the source-reading that produced each took longer and was confidently wrong.
  5. Withdrawal beats a third patch. Finding (c) had already been fixed twice. Two defects from one
    mechanism is the stop rule — the paragraph was deleted rather than corrected again, keeping only
    the source-verified sentence.
  6. Provenance is not verification. Claim (c) was faithful to ControlWatermarkOff's own
    [Description] string. Copying a claim from the code's own documentation is not checking it.
  7. A test that omits an optional key proves nothing about null handling — it passes under
    Newtonsoft-with-Ignore, Newtonsoft-default and System.Text.Json alike. Only an explicit
    "watermark": null distinguishes them.

Verification: python3 -m pytest scripts/tests → 956 passed, 2 skipped. decisions-validate: OK;
catalog regenerates byte-identical. check-doc-narrative --diff origin/main → 0 warnings. The
28-operation table re-derived by an independent $ref-resolving parser: zero missing, zero extra,
zero field mismatches. Both embedded snippets executed verbatim from ~/media-management and
~/server-management → 28 ops each. Boundary sweep by subject found no surviving stale claim.
Four rounds of cold, worktree-isolated adversarial review; Review-verdict: MERGEABLE @ 70526ce.
Full CI green on the merged head (the web/ CI-parity pre-push block was skipped as vacuous — the
diff contains no web/ and no .cs files).

Deferred:

  • ersatztv#837 — the measurable residual of finding (c): the documented "null → all off"
    behaviour depends on NullValueHandling.Ignore staying set, nothing records the dependency, and no
    test would catch it changing (a formatter swap turns a documented 200 into a 400).
  • Outside this repo, surfaced not fixed: ~/homelab-docs/Operations/Project Boundaries.md still
    contains a stale tree diagram — "ersatztv ├── Owns: ErsatzTV fork code, channel management" — which
    contradicts the corrected tables in the same file (lines 15-16, 86-87). homelab-docs is
    server-management's to change, so it was not touched here.

Docs updated: CLAUDE.md, docs/decisions/records/process/ersatztv-owns-code-not-operations.md
(new), docs/decisions/README.md (regenerated), .claude/skills/ersatztv/SKILL.md.

## Closing record **Outcome:** SHIPPED — PR #838 (squashed to `main`). `CLAUDE.md` → Project Boundaries now states the split, a decision record captures it (`process.ersatztv-owns-code-not-operations`), the canonical skill's banner names both symlinks, and the Scripted Schedule API table is folded into the skill. All five `## Done-when` boxes satisfied. **Root cause:** `CLAUDE.md` used one phrase — "channel/collection/schedule management" — for two different jobs. Writing a `ProgramSchedule` handler is fork code; creating a schedule on the live instance is an operation. Nothing in the wording distinguished them, so three files ended up with three answers about skill ownership and media-management's copy silently became a divergent fork describing a Blazor UI that #91b had removed. **Decisions/conventions changed:** Added `process.ersatztv-owns-code-not-operations` (new record, catalog regenerated). No supersessions. **Reusable knowledge:** 1. **`docs/endpoint-index.md` does NOT cover the scripted-schedule API.** The scripted spec is a *separate* OpenAPI document (`ErsatzTV/wwwroot/openapi/scripted-schedule.json`), and the index generator only walks `v1.json` — `grep -c scripted docs/endpoint-index.md` returns 0. Anyone answering "what endpoints exist" from the index will silently miss all 28 scripted operations. 2. **The spec ships as two files and they are not interchangeable.** `scripted-schedule.json` and `scripted-schedule-tagged.json` carry the same 28 paths and both return 200; they differ only in tag grouping, and Scalar's `/docs` renders the `-tagged` one. 3. **You cannot distinguish a wrong scripted base path from a stale `buildId` by probing.** Measured: unauthenticated both forms are 401 (the api-key filter precedes routing); authenticated with a non-live build id, both are 404 — including the *correct* path. Read the route in `ScriptedScheduleController.cs`; do not infer it from a status code. 4. **The big one — three claims about runtime behaviour were asserted from source-reading and every one was refuted by execution**, across four review rounds: (a) "only the un-tagged spec is fetchable" (all three return 200); (b) "a wrong path only becomes visible once the key is right" (the correct path 404s too); (c) an NRE warning built on `System.Text.Json` semantics, when these controllers use Newtonsoft (`Startup.cs:506-512`, `NullValueHandling.Ignore`) and MVC's implicit `[Required]` rejects an explicit null at model validation anyway — **400, never a 500**. Each was individually plausible and individually wrong. The cheap probe that settled each one took under two minutes; the source-reading that produced each took longer and was confidently wrong. 5. **Withdrawal beats a third patch.** Finding (c) had already been fixed twice. Two defects from one mechanism is the stop rule — the paragraph was deleted rather than corrected again, keeping only the source-verified sentence. 6. **Provenance is not verification.** Claim (c) was faithful to `ControlWatermarkOff`'s own `[Description]` string. Copying a claim from the code's own documentation is not checking it. 7. **A test that omits an optional key proves nothing** about null handling — it passes under Newtonsoft-with-Ignore, Newtonsoft-default and System.Text.Json alike. Only an explicit `"watermark": null` distinguishes them. **Verification:** `python3 -m pytest scripts/tests` → 956 passed, 2 skipped. `decisions-validate: OK`; catalog regenerates byte-identical. `check-doc-narrative --diff origin/main` → 0 warnings. The 28-operation table re-derived by an independent `$ref`-resolving parser: zero missing, zero extra, zero field mismatches. Both embedded snippets executed verbatim from `~/media-management` and `~/server-management` → 28 ops each. Boundary sweep by subject found no surviving stale claim. Four rounds of cold, worktree-isolated adversarial review; `Review-verdict: MERGEABLE @ 70526ce`. Full CI green on the merged head (the `web/` CI-parity pre-push block was skipped as vacuous — the diff contains no `web/` and no `.cs` files). **Deferred:** - **ersatztv#837** — the measurable residual of finding (c): the documented "null → all off" behaviour depends on `NullValueHandling.Ignore` staying set, nothing records the dependency, and no test would catch it changing (a formatter swap turns a documented 200 into a 400). - **Outside this repo, surfaced not fixed:** `~/homelab-docs/Operations/Project Boundaries.md` still contains a stale tree diagram — *"ersatztv ├── Owns: ErsatzTV fork code, channel management"* — which contradicts the corrected tables in the same file (lines 15-16, 86-87). homelab-docs is server-management's to change, so it was not touched here. **Docs updated:** `CLAUDE.md`, `docs/decisions/records/process/ersatztv-owns-code-not-operations.md` (new), `docs/decisions/README.md` (regenerated), `.claude/skills/ersatztv/SKILL.md`.
timothy removed the in-progress label 2026-08-26 19:30:00 +02:00
Author
Owner

Follow-up (vault): the deferred item in the closing record above is now also fixed —
homelab-docs cf27724. ~/homelab-docs/Operations/Project Boundaries.md contained an ownership tree
still reading "ersatztv — Owns: ErsatzTV fork code, channel management", contradicting the tables in
the same file that commit 1512972 had already corrected for this split. The tree now matches, and
"media-management (once ready)" is dropped since the repo exists with issues enabled.

Nothing deferred remains for #755 except ersatztv#837, which is tracked separately.

**Follow-up (vault):** the deferred item in the closing record above is now also fixed — homelab-docs `cf27724`. `~/homelab-docs/Operations/Project Boundaries.md` contained an ownership tree still reading *"ersatztv — Owns: ErsatzTV fork code, channel management"*, contradicting the tables in the same file that commit `1512972` had already corrected for this split. The tree now matches, and "media-management (once ready)" is dropped since the repo exists with issues enabled. Nothing deferred remains for #755 except ersatztv#837, which is tracked separately.
Sign in to join this conversation.
1 Participants
Notifications
Due Date
No due date set.
Dependencies

No dependencies set.

Reference: timothy/ersatztv#755