Skip to content

Add GitHub Copilot CLI as a supported harness #25

Open

Description

@Cecilia-BW

Feature request

backpass currently supports 6 harnesses (claude, codex, pi, opencode, grok, cursor-cli) via adapters in src/discovery/adapters/. GitHub Copilot CLI is not among them, so backpass scan/backpass report "no transcripts associated with this repo" even when Copilot CLI sessions exist for that repo.

Request: add a copilot.js adapter under src/discovery/adapters/ so Copilot CLI session transcripts can be discovered and fed into the loss/gradient-descent pipeline, same as the existing harnesses.

Where Copilot CLI stores transcripts

Session state is stored locally per-session under:

~/.copilot/session-state/<session-id>/

Each session folder contains checkpoints (markdown summaries) and a files/ subfolder for persistent artifacts. It would be great to get guidance on:

  • Whether there's a canonical index/manifest mapping session-id -> cwd (for Tier 1 repo association) and/or recorded git remote info (for Tier 2), similar to how codex/grok record cwd/git_remotes in their session files.
  • The exact JSON/JSONL shape of a session-state folder, so an adapter can be written to match the shared.js distillation contract used by the other adapters.

Happy to help test against a real repo with Copilot CLI history if that's useful.

Activity

  1. kunchenguid commented on Aug 23, 2026

    @kunchenguid
    Owner

    Speaking as Kun's firstmate.

    VISION verdict (per rule):

    • Evidence is the only currency: aligns if the adapter reads real turns, not checkpoints. GitHub's own docs say each CLI session under ~/.copilot/session-state/<id>/ includes events.jsonl (complete event log) plus workspace artifacts. Checkpoints/markdown summaries are not evidence and must not be the source.
    • The human owns the weights: aligns (discovery only; no write-path change).
    • Nothing the model says is taken on faith: aligns if the adapter has a pinned fixture and measures the real store, not a README row.
    • The budget is the constraint: cannot tell / n/a (no memory-file growth).
    • It reads what you already have: aligns, with the bar from VISION. A harness qualifies when it records real session transcripts. A store holding only a model's summary is not evidence. Supporting one means a pinned fixture and a fail-soft adapter. More harnesses are welcome if a new one cannot destabilise the ones already working. workspace.yaml / session-store.db can supply cwd/repo association; any association that cannot be made deterministically stays best-effort and opt-in.
    • Failure is loud and named: aligns if a missing or drifted store warns and is skipped.
    • Scope: aligns (this is discovery, not a new writer, CI, or provider).

    Decision: in scope. Labeling ready-for-pr.
    Adapter must: read events.jsonl (or an equivalent verbatim event stream), fail-soft, ship a pinned fixture, and not treat checkpoints as evidence. Do not guess a session-id→cwd map if it is not deterministic.

  2. thecrockster commented on Aug 27, 2026

    @thecrockster

    It would be great if Copilot CLI support was added 👍

  3. kunchenguid commented on Aug 27, 2026

    @kunchenguid
    Owner

    Speaking as Kun's firstmate.

    Still-real pass (FM-BP-CRON). Newer activity since stamp 2026-08-23T23:16:42Z: community +1 from thecrockster (2026-08-27T10:47:29Z). No implementation landed; no open PR helps this issue.

    Verified on latest main: 8466e65 (2026-08-27 08:14 UTC / 01:14 AM PT).

    • adapters dir: claude, codex, cursor-cli, cursor-ide, grok, hermes, opencode, pi, shared, sqlite; no copilot.js
    • discovery index ADAPTERS does not import Copilot
    • config ALL_HARNESSES excludes copilot
    • README harness table has no Copilot CLI row
    • no open or closed PR titled or searching for copilot; open PRs 51/36/29/21 unrelated

    VISION per-rule verdict (discovery adapter feature; evidence: issue body, prior triage, main tree, Copilot CLI chronicle/events.jsonl docs):

    1. Evidence is the only currency — aligns if adapter reads verbatim turns from events.jsonl (or equivalent), not checkpoint/markdown summaries. H-5 rejected adapter-over-summaries; real transcript store qualifies.
    2. The human owns the weights — aligns (discovery-only; no write-path change).
    3. Nothing the model says is taken on faith — aligns if pinned fixture measures the real store, not a README row alone.
    4. The budget is the constraint — n/a (no memory-file growth).
    5. It reads what you already have and owns nothing — aligns. Local ~/.copilot/session-state/; no upload; no backpass-held key. Bar: real transcripts + pinned fixture + fail-soft; more harnesses welcome if they cannot destabilise existing ones. workspace.yaml / session-store.db may supply cwd/repo association; non-deterministic association stays best-effort and opt-in; accuracy over coverage.
    6. Failure is loud and named — aligns if missing/drifted store warns and is skipped.
    7. Scope — aligns (discovery adapter, not reviewer/CI/linter/harness/provider; does not train ~/.claude/CLAUDE.md or widen writes).

    Decision: still in scope, still unhelped. Keeping ready-for-pr, leaving open. No implementation from this triage.

    Adapter bar unchanged: read events.jsonl (verbatim event stream), fail-soft, pinned fixture, never treat checkpoints as evidence, do not invent a session-id to cwd map when it is not deterministic.

  4. kunchenguid commented on Sep 10, 2026

    @kunchenguid
    Owner

    Speaking as Kun's firstmate.

    Still-real pass (scheduled wake / stale-restamp). Newer activity since stamp 2026-08-27T11:30:30Z: none on this issue. No implementation landed; no open PR helps this issue.

    Verified on latest main: dc4855f (2026-09-09 11:43:49 -0700 UTC / 2026-09-09 11:43 PDT).

    VISION per-rule verdict (discovery adapter feature; evidence: issue body, prior triage, main tree, Copilot CLI events.jsonl / session-state docs from prior pass):

    1. Evidence is the only currency — aligns if adapter reads verbatim turns from events.jsonl (or equivalent), not checkpoint/markdown summaries. A store of real session transcripts qualifies; summaries do not.
    2. The human owns the weights — aligns (discovery-only; no write-path change).
    3. Nothing the model says is taken on faith — aligns if a pinned fixture measures the real store, not a README row alone.
    4. The budget is the constraint — n/a (no memory-file growth).
    5. It reads what you already have and owns nothing — aligns. Local ~/.copilot/session-state/; no upload; no backpass-held key. Bar: real transcripts + pinned fixture + fail-soft; more harnesses welcome if they cannot destabilise existing ones. workspace.yaml / session-store.db may supply cwd/repo association; non-deterministic association stays best-effort and opt-in; accuracy over coverage.
    6. Failure is loud and named — aligns if missing/drifted store warns and is skipped.
    7. Scope — aligns (discovery adapter, not reviewer/CI/linter/harness/provider; does not train ~/.claude/CLAUDE.md or widen writes).

    Contract-class: new-default (once shipped, expected to join ALL_HARNESSES like peer discovery adapters); non-deterministic session→repo association remains opt-in/best-effort per VISION. Not restore (feature never landed).

    Decision: still in scope, still unhelped. Keeping ready-for-pr, leaving open. No implementation from this triage. No author ping.

    Adapter bar unchanged: read events.jsonl (verbatim event stream), fail-soft, pinned fixture, never treat checkpoints as evidence, do not invent a session-id→cwd map when it is not deterministic.

  5. kunchenguid commented on Sep 25, 2026

    @kunchenguid
    Owner

    Speaking as Kun's firstmate.

    Still-real pass (scheduled wake / stale-restamp / FM-BP-CRON). Newer activity since stamp 2026-09-10T19:45:22Z: none on this issue. No implementation landed; no open PR helps this issue.

    Verified on latest main: de6a9aa (chore(main): release backpass 0.1.27, 2026-09-24).

    • discovery adapters: claude, codex, cursor-cli, cursor-ide, grok, hermes, opencode, pi, shared, sqlite — no copilot.js
    • ALL_HARNESSES = claude, codex, pi, opencode, grok, cursor, hermes (no copilot); OPT_IN_HARNESSES = cursor-ide only
    • no open PR implements a Copilot CLI discovery adapter (search still only hits unrelated acpx copilot agent subcommand history)

    VISION per-rule verdict (discovery adapter feature; evidence: issue body, prior triage, main tree):

    1. Evidence is the only currency — aligns if adapter reads verbatim turns from events.jsonl (or equivalent), not checkpoint/markdown summaries.
    2. The human owns the weights — aligns (discovery-only; no write-path change).
    3. Nothing the model says is taken on faith — aligns if a pinned fixture measures the real store, not a README row alone.
    4. The budget is the constraint — n/a (no memory-file growth).
    5. It reads what you already have and owns nothing — aligns. Local ~/.copilot/session-state/; no upload; no backpass-held key. Bar: real transcripts + pinned fixture + fail-soft; more harnesses welcome if they cannot destabilise existing ones. Non-deterministic session→repo association stays best-effort and opt-in; accuracy over coverage.
    6. Failure is loud and named — aligns if missing/drifted store warns and is skipped.
    7. Scope — aligns (discovery adapter, not reviewer/CI/linter/harness/provider).

    Contract-class: new-default (once shipped, expected to join ALL_HARNESSES like peer discovery adapters); non-deterministic association remains opt-in/best-effort. Not restore (feature never landed).

    Decision: still in scope, still unhelped. Keeping ready-for-pr, leaving open. No implementation from this triage. No author ping.

    Adapter bar unchanged: read events.jsonl (verbatim event stream), fail-soft, pinned fixture, never treat checkpoints as evidence, do not invent a session-id→cwd map when it is not deterministic.

  6. kunchenguid commented on Oct 9, 2026

    @kunchenguid
    Owner

    Speaking as Kun's firstmate:

    Still-real pass (scheduled wake / stale-restamp / FM-BP-CRON). Newer activity since stamp 2026-09-25T04:32:45Z: none on this issue. No implementation landed; no open PR helps this issue.

    Verified on latest main: 0268201 (chore(main): release backpass 0.1.32).

    • discovery adapters: claude, codex, cursor-cli, cursor-ide, grok, hermes, opencode, pi, shared, sqlite — no copilot.js
    • ALL_HARNESSES = claude, codex, pi, opencode, grok, cursor, hermes (no copilot); OPT_IN_HARNESSES = cursor-ide only
    • discovery index ADAPTERS does not import Copilot
    • no open PR implements a Copilot CLI discovery adapter (search still only hits unrelated acpx copilot agent subcommand history)

    VISION per-rule verdict (discovery adapter feature; evidence: issue body, prior triage, main tree):

    1. Evidence is the only currency — aligns if adapter reads verbatim turns from events.jsonl (or equivalent), not checkpoint/markdown summaries.
    2. The human owns the weights — aligns (discovery-only; no write-path change).
    3. Nothing the model says is taken on faith — aligns if a pinned fixture measures the real store, not a README row alone.
    4. The budget is the constraint — aligns (N/A; no memory-file growth).
    5. It reads what you already have and owns nothing — aligns. Local ~/.copilot/session-state/; no upload; no backpass-held key. Bar: real transcripts + pinned fixture + fail-soft; more harnesses welcome if they cannot destabilise existing ones. Non-deterministic session→repo association stays best-effort and opt-in; accuracy over coverage.
    6. Failure is loud and named — aligns if missing/drifted store warns and is skipped.
    7. Scope — aligns (discovery adapter, not reviewer/CI/linter/harness/provider).
    8. Resist widen-writes / lower two-session / default non-per-edit / easier-to-grow / hold key-server-transcripts / coverage-over-accuracy — aligns (none of those if adapter bar above is kept).

    Contract-class: new-default (once shipped, expected to join ALL_HARNESSES like peer discovery adapters); non-deterministic association remains opt-in/best-effort. Not restore (feature never landed).

    Decision: still in scope, still unhelped. Keeping ready-for-pr, leaving open. No implementation from this triage. No author ping.

    Adapter bar unchanged: read events.jsonl (verbatim event stream), fail-soft, pinned fixture, never treat checkpoints as evidence, do not invent a session-id→cwd map when it is not deterministic.

  7. kunchenguid commented on Oct 23, 2026

    @kunchenguid
    Owner

    Speaking as Kun's firstmate:

    Still-real pass. Nothing new on this issue since the 2026-10-09 stamp, no implementation has landed, and no open PR helps it. On latest main (3882b86) the discovery adapters are still claude, codex, cursor-cli, cursor-ide, grok, hermes, opencode, pi, shared, and sqlite, with no copilot.js.

    The earlier VISION verdict stands: every rule aligns, as long as the adapter reads verbatim turns from events.jsonl (not checkpoint summaries), ships a pinned fixture, fails soft when the store is missing or drifted, and keeps any non-deterministic session-to-repo association best-effort. Contract-class is new-default if it ships on by default, so a PR will go to Kun for that decision. It stays ready-for-pr, and a contribution is welcome.

  8. kunchenguid commented on Nov 6, 2026

    @kunchenguid
    Owner

    Speaking as Kun's firstmate:

    Still-real pass (scheduled wake / stale-restamp / FM-BP-CRON). Newer activity since stamp 2026-10-23T16:48:00Z: none on this issue. No implementation landed; no open PR helps this issue.

    Verified on latest main: c8d60c2 (chore(main): release backpass 0.1.44, 2026-11-02).

    VISION per-rule verdict (discovery adapter feature; evidence: issue body, prior triage, main tree, Copilot CLI ~/.copilot/session-state/ / events.jsonl docs from prior passes):

    1. Evidence is the only currency — aligns if adapter reads verbatim turns from events.jsonl (or equivalent), not checkpoint/markdown summaries. A store of real session transcripts qualifies; summaries do not.
    2. The human owns the weights — aligns (discovery-only; no write-path change).
    3. Nothing the model says is taken on faith — aligns if a pinned fixture measures the real store, not a README row alone.
    4. The budget is the constraint — aligns (N/A; no memory-file growth).
    5. It reads what you already have and owns nothing — aligns. Local ~/.copilot/session-state/; no upload; no backpass-held key. Bar: real transcripts + pinned fixture + fail-soft; more harnesses welcome if they cannot destabilise existing ones. Non-deterministic session→repo association stays best-effort and opt-in; accuracy over coverage.
    6. Failure is loud and named — aligns if missing/drifted store warns and is skipped.
    7. Scope — aligns (discovery adapter, not reviewer/CI/linter/harness/provider; does not train ~/.claude/CLAUDE.md or widen writes).
    8. Resist widen-writes / lower two-session / default non-per-edit / easier-to-grow / hold key-server-transcripts / coverage-over-accuracy — aligns (none of those if adapter bar above is kept).

    Contract-class: new-default (once shipped, expected to join ALL_HARNESSES like peer discovery adapters); non-deterministic association remains opt-in/best-effort. Not restore (feature never landed).

    Decision: still in scope, still unhelped. Keeping ready-for-pr, leaving open. No implementation from this triage. No author ping.

    Adapter bar unchanged: read events.jsonl (verbatim event stream), fail-soft, pinned fixture, never treat checkpoints as evidence, do not invent a session-id→cwd map when it is not deterministic.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    ready-for-prIssue is in scope and ready for a PR

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions