@azure-id/orc 2.0.2 → 2.1.0

package/mock-run/orc-budget.md

@@ -72,14 +72,11 @@ ORC · budget · TOKENS  (p50 → p90)
 
   band          model / count        input    cache-w    cache-r     output
   ────────────────────────────────────────────────────────────────────────────
-  [0,30)   ×2   haiku-4-5             1.2k       6.4k      11.8k       0.9k
-  [30,40)  ×1   sonnet-4-6-med        2.1k      14.0k      28.4k       2.2k
-  [40,55)  ×3   sonnet-4-6-high       7.8k      52.1k     104.0k       8.4k
-  [55,65)  ×2   sonnet-5-high         6.9k      48.3k      96.7k       9.1k
-  [65,70)  ×1   opus-4-7-med          4.4k      31.0k      62.5k       6.8k
-  [70,80)  ×1   opus-4-7-high         5.7k      40.2k      81.0k       9.3k
-  [80,90)  ×1   opus-4-8-high         7.1k      49.8k     101.4k      12.0k
-  [90,100] ×1   opus-5-high           8.0k      56.0k     114.0k      14.2k
+  [0,21)   ×1   sonnet-5-low          0.6k       3.2k       5.9k       0.4k
+  [21,31)  ×1   sonnet-5-med          0.6k       3.2k       5.9k       0.5k
+  [31,41)  ×1   sonnet-5-high         2.1k      14.0k      28.4k       2.2k
+  [41,90)  ×8   opus-5-low           31.9k     221.4k     445.6k      45.6k
+  [90,100] ×1   opus-5-med            8.0k      56.0k     114.0k      14.2k
   fixed roles   analyst · planner · reviewer · verifier · trace writer
                                      18.2k     121.0k     248.0k      26.4k
   ────────────────────────────────────────────────────────────────────────────
@@ -101,7 +98,7 @@ ORC · budget · the other views
 
 ORC · budget · CONTEXT RISK — 1 of 12 tasks
 
-  T12  backfill 1.4M rows   opus-5-high   est. peak 189k / 200k   95%   ⚠
+  T12  backfill 1.4M rows   opus-5-med    est. peak 189k / 200k   95%   ⚠
        what is in the slice:
          3 wiki docs                58k    ← orc-reference-config.md used 2/20
          code pattern (typescript)  11k
@@ -118,6 +115,11 @@ ORC · budget · CONTEXT RISK — 1 of 12 tasks
   Other 11 tasks: highest is T09 at 142k / 200k (71%). Fine.
 ```
 
+> The dollar figures in this walkthrough are illustrative. The token counts
+> add up. The dated price table sets the real dollars: Sonnet 5 is
+> $2 / $2.50 / $0.20 / $10 per MTok (input / cache write / cache read /
+> output), Opus 5.5 is $4 / $5 / $0.20 / $20, Haiku 4.5 is $1 / $5.
+
 ```
 ORC · budget · the same plan, on other lanes
 
@@ -143,17 +145,14 @@ ORC · budget · how sure am I
 
   band          past dispatches   confidence
   ─────────────────────────────────────────────
-  [0,30)              41          good
-  [30,40)             27          good
-  [40,55)             38          good
-  [55,65)             31          good
-  [65,70)             14          fair
-  [70,80)              9          fair
-  [80,90)              4          LOW — insufficient history
+  [0,21)              23          good
+  [21,31)             18          good
+  [31,41)             27          good
+  [41,90)             96          good
   [90,100]             2          LOW — insufficient history
 
-  Two bands have very little history. Their p90 is shaped by the band next
-  to them, not measured. Treat the top of the range as soft.
+  One band has very little history. Its p90 is shaped by the band next
+  to it, not measured. Treat the top of the range as soft.
 
   Cache-read has its own range and it is the widest part of the forecast —
   it depends on how much of your repo is already warm. Do not budget on the
@@ -229,13 +228,11 @@ BUDGET · actual vs forecast · run-orc-partial-refunds-100826-171450
 
   band        forecast p50 (weighted)   actual     diff      note
   ────────────────────────────────────────────────────────────────────────
-  [0,30)              9k                  9k       on target
-  [30,40)            21k                 20k       -5%
-  [40,55)            92k                 79k       -14%
-  [55,65)            86k                121k       +41%      T08 2 repair rounds
-  [65,70)            57k                 54k       -6%
-  [70,80)            74k                186k      +151%      T04 requeued twice
-  [80,90)            96k                101k       +5%
+  [0,21)              4k                  4k       on target
+  [21,31)             5k                  5k       on target
+  [31,41)            21k                 20k       -5%
+  [41,90)           405k                541k       +34%      T04 requeued twice (+151%)
+                                                             T08 2 repair rounds
   [90,100]            —                   —        not run   T12 blocked
   fixed roles       178k                187k       +5%
   ────────────────────────────────────────────────────────────────────────
@@ -257,7 +254,7 @@ BUDGET · actual vs forecast · run-orc-partial-refunds-100826-171450
 
     Same finding from three angles. Cheap fix: write the test for PACT-006.
 
-  Fed to /orc-retro. Band [70,80) now has 10 samples.
+  Fed to /orc-retro. Band [41,90) now has 104 samples.
 ```
 
 **Why both weighted and raw are reported:** raw is what fills a context window
@@ -308,7 +305,7 @@ was used in 2 of the last 20 runs.
 
 ## Confidence
 
-Bands `[80,90)` and `[90,100]` have fewer than 5 samples. Their p90 is derived,
+Band `[90,100]` has fewer than 5 samples. Its p90 is derived,
 not measured. `unattributed` = 2.6%.
 
 ## Assumptions
@@ -376,7 +373,6 @@ $ orc budget forecast plan.md --json
     }
   ],
   "bands": [
-    { "band": "[80,90)", "samples": 4, "confidence": "low", "reason": "insufficient history" },
     { "band": "[90,100]", "samples": 2, "confidence": "low", "reason": "insufficient history" }
   ],
   "exit": 2
@@ -431,9 +427,9 @@ BAND RATES · from 189 dispatches
 
   band          n    input   cache-w   cache-r   output   weighted p50
   ──────────────────────────────────────────────────────────────────────
-  [0,30)       41    0.6k     3.2k       5.9k     0.5k      4.4k
-  [40,55)      38    2.6k    17.4k      34.7k     2.8k     30.6k
-  [70,80)       9    5.7k    40.2k      81.0k     9.3k     74.0k   fair
+  [0,21)       23    0.6k     3.2k       5.9k     0.4k      4.8k
+  [31,41)      27    2.1k    14.0k      28.4k     2.2k     21.1k
+  [41,90)      96    4.0k    27.7k      55.7k     5.7k     43.0k
   [90,100]      2    8.0k    56.0k     114.0k    14.2k    103.0k   LOW
 ```
 
@@ -458,7 +454,7 @@ Intake · confirm before we start
   Forecast   1.42M raw / 652k weighted tokens
              21% → 32% of your 5-hour window
              $6.76 → $10.50 · 18 → 31 minutes
-             (from your own 31 runs; 2 bands have low confidence)
+             (from your own 31 runs; 1 band has low confidence)
 
   ⚠ T12 is forecast at 95% of its context window and will probably compact.
 
@@ -526,7 +522,7 @@ says so, then sharpens. That is a good shape for a product: it rewards staying.
 
 **It closes a loop with `/orc-retro`.** Forecast vs actual per band is exactly
 the signal retro needs to tune the score→model table. Today retro sees that a
-band retried a lot. Now it also sees that the band cost 151% more — and that
+band retried a lot. Now it also sees that one task in the band cost 151% more — and that
 the overage was mostly *output* tokens, which points straight at retries rather
 than at big context.
 

package/mock-run/orc-cli.md

@@ -1,200 +1,200 @@
-# Mock run — the `orc` terminal commands
-
-> Everything here costs **zero model tokens**. It is plain Node reading files.
-
-The CLI is not a side door to the skills. It owns the parts that must be
-deterministic: installing, settings, probes with exit codes, and reading back
-run state.
-
----
-
-## 1. Install and health
-
-```bash
-$ orc init
-```
-
-```
-Installed into /home/rina/shopcart/.claude
-  skills    30
-  commands  27
-  agents    40  (+ MODEL-MAPPING.md)
-  hooks     3   effort guard · statusline · behavior trace
-  manifest  .claude/orc/install-manifest.json
-
-Merged into .claude/settings.json (nothing of yours was replaced):
-  · PreToolUse effort guard
-  · statusline  (you had none — if you had, I would print the snippet instead)
-
-Next: orc onboarding first-run
-```
-
-```bash
-$ orc doctor
-```
-
-```
-orc doctor — /home/rina/shopcart/.claude
-
-  version      0.46.0 project · 0.44.0 GLOBAL ~/.claude
-               ⚠ the global install is older and can win skill resolution
-               fix: orc update --global
-  payload      ✓ every shipped file present
-  orphans      3 files ORC owns are no longer in the payload
-               fix: orc update --prune   (it will name all three first)
-  settings     ✓ guard wired · ✓ statusline wired
-  trace        ✓ pointer valid
-  diy          STALE — compiled before the last update
-
-3 findings                                                        (exit 1)
-```
-
----
-
-## 2. Settings, without spending tokens
-
-```bash
-$ orc config recommend
-```
-
-```
-Reading this repo (read-only)…
-  a real `npm test` script exists          → gates have something to check
-  CI is configured                         → this repo is shared
-  7 contributors in history                → coordination cost is real
-  a project wiki exists                    → grounding is already cheap
-
-Recommended profile: paranoid
-Apply with: orc config profile paranoid
-```
-
-```bash
-$ orc config set max_wave_tasks 4
-max_wave_tasks: 3 → 4      written to .claude/orc.config.yaml
-
-$ orc config set opus5_only true
-opus5_only: false → true
-  ⚠ this makes 4 keys inert while it is on:
-      fable5_enabled, fable5_effort, fable5_roles, rubric_bands_override
-    They are shown as shadowed in `orc config list`.
-```
-
----
-
-## 3. Probes — the exit code is the answer
-
-These are what the knowledge-gated lanes run before they start. They print one
-line and set an exit code, so nothing has to guess.
-
-```bash
-$ orc pattern status express   ;  echo "exit $?"
-express   cached   2026-08-12
-exit 0
-
-$ orc wiki status
-wiki   AGING   17 docs · worst doc 24 commits behind · edges 10 / 30
-
-$ orc wiki impact              ;  echo "exit $?"
-4 of 17 docs TOUCHED, 0 STRUCTURAL → a delta refresh is enough
-exit 2
-
-$ orc diy status               ;  echo "exit $?"
-UNCONFIGURED — no .claude/orc-diy.config.yaml
-exit 1
-```
-
----
-
-## 4. Reading run state back
-
-```bash
-$ orc run list
-```
-
-```
-  slug                     lane      state        when
-  refund-visibility        orc       waiting      2 hours ago
-  order-notes              orc       finished     yesterday
-  badge-fix                quick     finished     yesterday
-  csv-export               diy       incomplete   4 days ago
-
-`waiting` = RESUME.md exists. `incomplete` = no finish and no resume file —
-that is all the disk proves, so that is all this says.
-```
-
-```bash
-$ orc resume
-```
-
-```
-1 run is waiting:
-  1  refund-visibility   /orc · phase 3 · wave 2 of 3
-
-Pick one:  > 1
-
-Paste this into a fresh Claude Code session (already copied to your clipboard):
-
-  | Continue ORC run `refund-visibility`.
-  | Read .claude/orc/run/refund-visibility/state-of-play.md, then checkpoint.json.
-  | Resume from the checkpoint phase/wave. Do not re-plan. Do not redo done tasks.
-```
-
-```bash
-$ orc stats --since 2026-07-01
-```
-
-```
-Lanes          runs        Agents                          dispatches
-  orc            12          orc-executor-sonnet-4-6-med       41
-  quick          31          orc-executor-sonnet-5-high        22
-  mini            9          orc-planner-opus-5-med            21
-  wiki            3          orc-reviewer-opus-5-med           12
-  fast            2          orc-wiki-scanner-opus-4-8-high     9
-
-Downgrades caught: 2   (a subagent answered on a lower model than its name)
-Counted from trace filenames. No model ran to produce this.
-```
-
----
-
-## 5. Mocked runs, from the terminal
-
-```bash
-$ orc mock-run list
-```
-
-```
-mocked runs — 24 documents that ship with ORC
-
-Start here
-  the-example-project        the shopcart project every mock uses
-  a-normal-day               all six new lanes inside one ordinary run
-Build a change
-  orc                        the full pipeline, start to ship
-  orc-ultra                  advisor + three judges
-  …
-
-Read one:  orc mock-run show orc-pact
-Or open the panel:  orc ui   ▸ Mocked Skill Use
-```
-
----
-
-## 6. What to notice
-
-- **`update` and `upgrade` are different.** `update` re-copies what is already
-  in this package (offline). `upgrade` fetches the newest package first, then
-  copies. Your `.claude/orc.config.yaml` survives both.
-- **Every read command speaks `--json`** — exactly one object on stdout and the
-  same exit code the human output would use. That is what `orc ui` is built on.
-- **Empty is an answer.** A command with no results still prints its object and
-  says so, rather than looking broken.
-- **`orc onboarding`** is the whole walkthrough in the terminal, so you never
-  need this repository open to get started.
-
----
-
-## 7. Related
-
-- The same things with buttons: [`orc ui`](orc-ui.md)
-- Full command list: `orc --help`
+# Mock run — the `orc` terminal commands
+
+> Everything here costs **zero model tokens**. It is plain Node reading files.
+
+The CLI is not a side door to the skills. It owns the parts that must be
+deterministic: installing, settings, probes with exit codes, and reading back
+run state.
+
+---
+
+## 1. Install and health
+
+```bash
+$ orc init
+```
+
+```
+Installed into /home/rina/shopcart/.claude
+  skills    30
+  commands  27
+  agents    40  (+ MODEL-MAPPING.md)
+  hooks     3   effort guard · statusline · behavior trace
+  manifest  .claude/orc/install-manifest.json
+
+Merged into .claude/settings.json (nothing of yours was replaced):
+  · PreToolUse effort guard
+  · statusline  (you had none — if you had, I would print the snippet instead)
+
+Next: orc onboarding first-run
+```
+
+```bash
+$ orc doctor
+```
+
+```
+orc doctor — /home/rina/shopcart/.claude
+
+  version      0.46.0 project · 0.44.0 GLOBAL ~/.claude
+               ⚠ the global install is older and can win skill resolution
+               fix: orc update --global
+  payload      ✓ every shipped file present
+  orphans      3 files ORC owns are no longer in the payload
+               fix: orc update --prune   (it will name all three first)
+  settings     ✓ guard wired · ✓ statusline wired
+  trace        ✓ pointer valid
+  diy          STALE — compiled before the last update
+
+3 findings                                                        (exit 1)
+```
+
+---
+
+## 2. Settings, without spending tokens
+
+```bash
+$ orc config recommend
+```
+
+```
+Reading this repo (read-only)…
+  a real `npm test` script exists          → gates have something to check
+  CI is configured                         → this repo is shared
+  7 contributors in history                → coordination cost is real
+  a project wiki exists                    → grounding is already cheap
+
+Recommended profile: paranoid
+Apply with: orc config profile paranoid
+```
+
+```bash
+$ orc config set max_wave_tasks 4
+max_wave_tasks: 3 → 4      written to .claude/orc.config.yaml
+
+$ orc config set opus5_only true
+opus5_only: false → true
+  ⚠ this makes 4 keys inert while it is on:
+      fable5_enabled, fable5_effort, fable5_roles, rubric_bands_override
+    They are shown as shadowed in `orc config list`.
+```
+
+---
+
+## 3. Probes — the exit code is the answer
+
+These are what the knowledge-gated lanes run before they start. They print one
+line and set an exit code, so nothing has to guess.
+
+```bash
+$ orc pattern status express   ;  echo "exit $?"
+express   cached   2026-08-12
+exit 0
+
+$ orc wiki status
+wiki   AGING   17 docs · worst doc 24 commits behind · edges 10 / 30
+
+$ orc wiki impact              ;  echo "exit $?"
+4 of 17 docs TOUCHED, 0 STRUCTURAL → a delta refresh is enough
+exit 2
+
+$ orc diy status               ;  echo "exit $?"
+UNCONFIGURED — no .claude/orc-diy.config.yaml
+exit 1
+```
+
+---
+
+## 4. Reading run state back
+
+```bash
+$ orc run list
+```
+
+```
+  slug                     lane      state        when
+  refund-visibility        orc       waiting      2 hours ago
+  order-notes              orc       finished     yesterday
+  badge-fix                quick     finished     yesterday
+  csv-export               diy       incomplete   4 days ago
+
+`waiting` = RESUME.md exists. `incomplete` = no finish and no resume file —
+that is all the disk proves, so that is all this says.
+```
+
+```bash
+$ orc resume
+```
+
+```
+1 run is waiting:
+  1  refund-visibility   /orc · phase 3 · wave 2 of 3
+
+Pick one:  > 1
+
+Paste this into a fresh Claude Code session (already copied to your clipboard):
+
+  | Continue ORC run `refund-visibility`.
+  | Read .claude/orc/run/refund-visibility/state-of-play.md, then checkpoint.json.
+  | Resume from the checkpoint phase/wave. Do not re-plan. Do not redo done tasks.
+```
+
+```bash
+$ orc stats --since 2026-07-01
+```
+
+```
+Lanes          runs        Agents                          dispatches
+  orc            12          orc-executor-sonnet-4-6-med       41
+  quick          31          orc-executor-sonnet-5-high        22
+  mini            9          orc-planner-opus-5-med            21
+  wiki            3          orc-reviewer-opus-5-low           12
+  fast            2          orc-wiki-scanner-opus-5-low     9
+
+Downgrades caught: 2   (a subagent answered on a lower model than its name)
+Counted from trace filenames. No model ran to produce this.
+```
+
+---
+
+## 5. Mocked runs, from the terminal
+
+```bash
+$ orc mock-run list
+```
+
+```
+mocked runs — 24 documents that ship with ORC
+
+Start here
+  the-example-project        the shopcart project every mock uses
+  a-normal-day               all six new lanes inside one ordinary run
+Build a change
+  orc                        the full pipeline, start to ship
+  orc-ultra                  advisor + three judges
+  …
+
+Read one:  orc mock-run show orc-pact
+Or open the panel:  orc ui   ▸ Mocked Skill Use
+```
+
+---
+
+## 6. What to notice
+
+- **`update` and `upgrade` are different.** `update` re-copies what is already
+  in this package (offline). `upgrade` fetches the newest package first, then
+  copies. Your `.claude/orc.config.yaml` survives both.
+- **Every read command speaks `--json`** — exactly one object on stdout and the
+  same exit code the human output would use. That is what `orc ui` is built on.
+- **Empty is an answer.** A command with no results still prints its object and
+  says so, rather than looking broken.
+- **`orc onboarding`** is the whole walkthrough in the terminal, so you never
+  need this repository open to get started.
+
+---
+
+## 7. Related
+
+- The same things with buttons: [`orc ui`](orc-ui.md)
+- Full command list: `orc --help`