Skip to content

♻️ Retire the MCP browser-cookie read bridge - #95

Merged
arach merged 4 commits into
core-splitfrom
lane/c-e3-herdr-runtime-read
Oct 23, 2026
Merged

arach merged 4 commits into
core-splitfrom
lane/c-e3-herdr-runtime-read

Conversation

@arach

@arach arach commented Oct 22, 2026 •

Copy link
Copy Markdown
Owner

Latest rebase and checkpoint accounting

Head 870187c8a6440b1371707c9e805de2ba8a9249b5, rebased onto core-split 4f328d41d72fb3d05105a77fb427228e310aa6aa after #107. The two source-inventory conflicts preserve #100's readiness tombstones and negative tests, alongside #95's adapter retirement. The reader, async extraction/body pipeline and all three reviewed policy transitions are otherwise unchanged.

Appended the named 175 -> 174 record in scripts/extension-channel-fence-inventory.json, with reviewed input f3bf9527705b102f2b46f68ea3cb53af83c7a710:

  • retire packages/cli/src/cli/commands/web.ts / spawnSync;
  • retire packages/cli/src/cli/commands/web.test.ts / spawnSync;
  • add packages/cli/src/cli/commands/fixtures/web-stdio.ts / Bun.spawnSync.

All complete-policy digests match the supplied reviewed transition; no policy bytes changed in this rebase. The pinned baseline and the separate #100 record remain unchanged. This is an appended accounting record, not a baseline regeneration or a disk recount.

Fresh local evidence at this exact head (macOS arm64)

Check Bun 1.3.14 Bun 1.3.13
Canonical sync-exec fence PASS, 174 pairs PASS, 174 pairs
Checkpoint helper suite 7 pass / 0 fail 7 pass / 0 fail
CLI private-source inventory negatives 11 pass / 0 fail 11 pass / 0 fail
Full extension-channel checkpoint PASS, final expected inventory 174 Correct unsupported-coordinate refusal, exit 1

The checkpoint intentionally qualifies only the pinned Bun 1.3.14 revision on supported hosts. Its unchanged predicate refuses 1.3.13 before any native cohort; this is not a passing qualification claim for 1.3.13. The 1.3.14 full run completes the cold witness, all 28 warm matrix cases, 105 cohort cases, 25 runner cases, package test and expected-inventory fence, with no positive skips.

Tests initially waited below the 3-GiB disk floor. Space later recovered; every executed command passed its pre-test guard and was disk-monitored. Hosted Ubuntu checkpoint evidence remains outstanding; the point-owned hosted matrix also supplies macos-14 confirmation. No CI was dispatched. Earlier revision evidence below remains historical, not relabeled to this rebase.


One concern: retire the retained MCP browser-cookie bridge

Base c9ea8c9beebacfc4d2519648be71d62b34a111b1; head ef1f4920be32507ea117ad0a4640f06416fe14bb.

The final production caller of cli/web-api.ts is herdr_workspaces. Use the
reviewed runtime /v1/terminal-hosts/herdr/workspaces projection instead, then
remove the adapter, its obsolete web-auth tests and its exact temporary private
inventory exception. Add the retired source to the inventory's negative list.

  • One selected local broker GET; Unix selection has no HTTP fallback. No web
    bearer/cookie/bootstrap, local Herdr shell, app start, retry or alternate source.
  • Eight-second fetch/body signal, 2 MiB response bound, strict versioned/nested
    validation, exact session correlation, live/saved consistency and safe HTTP
    status/code errors. Redirects/ambient credentials disabled.
  • Preserve ranked digest/query/limit behavior and panel/session provenance.
    Keep not-scanned/unavailable distinct from available-empty; partial empty
    observations remain incomplete, never absence. The actual runtime's 256
    distinct-directory groups with only 40 emitted panes are accepted.
  • Maintained interface documentation: docs/herdr-workspaces-mcp.md.

API review required

Additive MCP result change: runtimeRead preserves broker schema/node/time,
availability, reason, count, limits and per-session availability; readError
preserves safe HTTP status/code. The backend is now the machine-local broker,
not a configured private web origin. An older/refusing broker is unavailable;
there is deliberately no legacy web fallback. Tool name/inputs and existing
result fields remain. No runtime/protocol API, package export, dependency or
new host-control/answer capability. Please route the MCP behavior to Fable.

Measurement and selection

Same five roots and AST algorithm, Git blobs at each exact revision:

Metric e823cce c9ea8c9 This head
Non-CLI desktop values 84 36 36
Local value closure 189 209 209
Expanded values 582 625 625
Expanded values plus types 621 670 670
Unresolved relative/alias value edges 0 0 0

M0 removes 49 historical desktop paths from that category; one explicit private
pair HTTP proxy is added, for a net 48 reduction. Public API/client growth makes
total file counts larger, not evidence of independence. Classification is
transported through M0, never relabeled by folder. This PR removes one retirement
adapter: directed P+D edge occurrences 71 -> 70, distinct P/D targets 39 -> 38;
strict P remains 68, and local private files remain 38. The new bounded
public reader replaces one file, so raw closure totals do not shrink.

The largest remaining family is pair: 34 desktop + six web files. Its removal
is gated on private providers, qualified delivery, preseed and the phone/npm
product decision. Setup/doctor still need core-only readiness/custody. This PR
finishes the largest remaining approved retained-read boundary without flipping
any private built-in or touching those gates. It is not whole-CLI independence.

Exact-base/head local verification

Fresh disposable HOME/TMP/state per command, both full CLI suites sequentially;
all new temporary homes removed. macOS arm64.

Gate Bun 1.3.14 base -> head Bun 1.3.13 base -> head
Focused reader coverage 16/0 -> 29/0 16/0 -> 29/0
Full CLI source 1309/0 -> 1322/0 1309/0 -> 1322/0
CLI typecheck pass -> pass pass -> pass
Whole sync-exec fence 173 pairs -> 173 pairs 173 pairs -> 173 pairs

Node 22.22.2 inventory checker: 9/0, including retired-source reappearance.
Both Buns compile the affected reader: eight workspace inputs, no desktop/web or
retired adapter, and no browser-cookie/bootstrap code. This is a scoped JS proof,
not a full npm/native candidate or a claim that all app bundles have left npm.

All three deliberate defects fail their own targeted assertion on both Buns:
remove Unix selection; fabricate availability; remove requested-session binding.
No import error/watchdog counts as a valid control; mutations restored.

Initial dependency/fixture development failures and the receipt-classifier
correction are preserved in the evidence ledger. No inherited failure waiver,
removed unrelated coverage, assertion relaxation or timeout increase. The four
removed tests cover the deleted adapter's browser authentication; transport and
refusal coverage now exercises the broker instead.

Private raw evidence and same-instrument census:
/Users/art/.scout/reports/core-split-c-e3-next-20261022/REPORT.md and
MEASUREMENT.md, with per-version commands/logs/hashes, controls and bundler metadata.

Draft for review. No CI dispatch, native build, live service/app operation,
provider flip, pin/cutover activation, publication or merge.

@arach

arach commented Oct 22, 2026

Copy link
Copy Markdown
Owner Author

Rebased onto 40fa893bf97365c05a13c7bc009a2a300164506e and added the download-consumer deadline at head 84864845bf26e6512cd57c800857601cd7c74374.

The cancelled run completed all 29 Herdr reader/adapter tests. Its last web pass precedes an upgraded CLI diagnoses the old Full client and recovers with the saved key; no stack identifies the exact awaited subcall. The concrete gap in that path was a five-minute signal passed to fetch but not the body/file pipeline. One signal now bounds both, with stalled-body and stalled-upgrade regressions; checksum, old-client/key preservation, atomic staging and the existing tar deadline remain.

On both Buns, on macOS and Ubuntu 24.04 x86_64: focused 48/0, CLI typecheck pass, fence 173. Full CLI: macOS 1324/0; Linux 1323 pass, one unchanged Darwin-only skip, zero failures. Removing only the consumer signal causes a normal targeted assertion failure on all four host/Bun combinations.

Important limit: the original #95 full CLI also passed on the Linux VM (1321 pass, one platform skip, zero failures), as did 25 original web-file repeats and 200 instrumented reruns. The original web blob was identical to the base. The natural hosted trigger is therefore still unconfirmed; this is a measured deadline correction, not proof of a #95-only or Linux-only root cause. No CI was dispatched. Point/operator owns any hosted rerun.

@arach
arach force-pushed the lane/c-e3-herdr-runtime-read branch from f3bf952 to 870187c Compare October 23, 2026 02:17
@arach
arach marked this pull request as ready for review October 23, 2026 03:12
@arach
arach merged commit 8130234 into core-split Oct 23, 2026
4 checks passed
@arach
arach deleted the lane/c-e3-herdr-runtime-read branch November 20, 2026 22:20
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant